NBER WORKING PAPER SERIES

PRODUCT DIFFERENTIATION, MULTI-PRODUCT FIRMS AND ESTIMATING THE IMPACT OF TRADE LIBERALIZATION ON PRODUCTIVITY Jan De Loecker Working Paper 13155 http://www.nber.org/papers/w13155

NATIONAL BUREAU OF ECONOMIC RESEARCH 1050 Massachusetts Avenue Cambridge, MA 02138 June 2007

An earlier version of this paper was circulated under the title "Product Differentiation, Multi-Product Firms and Structural Estimation of Productivity". I am especially grateful to Joep Konings, Marc Melitz and Ariel Pakes for comments and suggestions. I would also like to thank Dan Ackerberg, Steve Berry, David Blackburn, Mariana Colacelli, Bronwyn Hall, Amil Petrin, Frank Verboven, Patrick Van Cayseele and Hylke Vandenbussche for discussions on earlier versions. This paper has benefited from seminar participants at KU Leuven, Princeton, NYU, U of Toronto, UT Austin, UCSC, IESE, Federal Board of Governors, LSE, WZB, NY Fed and conference participants at IIOC 2005, CEPR Applied IO Workshop 2005, NBER Summer Institute Productivity Workshop 2005, Midwest Meeting in International Economics 2005, NBER Productivity Workshop 2006. Furthermore, I would like to thank Roberta Adinolfi of EURATEX, Sylvie Groeninck of FEBELTEX, Yannick Hamon and Ignacio Iruarruzaga Diez at the European Commission for providing me with the relevant conversion tables to construct the quota data set. The views expressed herein are those of the author(s) and do not necessarily reflect the views of the National Bureau of Economic Research. 2007 by Jan De Loecker. All rights reserved. Short sections of text, not to exceed two paragraphs, may be quoted without explicit permission provided that full credit, including notice, is given to the source.

Product Differentiation, Multi-product Firms and Estimating the Impact of Trade Liberalization on Productivity Jan De Loecker NBER Working Paper No. 13155 June 2007 JEL No. F13,L11 ABSTRACT In this paper I analyze the productivity gains from trade liberalization in the Belgian textile industry. So far, empirical research has established a strong relationship between opening up to trade and productivity, relying almost entirely on deflated sales to proxy for output in the production function. The latter implies that the resulting productivity estimates still capture price and demand shocks which are most likely to be correlated with the change in the operating environment, which invalidate the evaluation of the welfare implications. In order to get at the true productivity gains I propose a simple methodology to estimate a production function controlling for unobserved prices by introducing an explicit demand system. I combine a unique data set containing matched plant-level and product-level information with detailed product-level quota protection information to recover estimates for productivity as well as parameters of the demand side (markups). I find that when correcting for unobserved prices and demand shocks, the estimated productivity gains from relaxing protection are only half (from 6 to only 3 percent) of those obtained with standard techniques. Jan De Loecker Stern School of Business New York University 44 West Fourth Street New York, NY 10012 and NBER jdl16@nyu.edu

Introduction

It is by now a well documented nding that periods of major changes in the competitive environment of rms - like trade liberalization - are associated with productivity gains and that rms engaged in international trade (through export or FDI) are more productive.1 The productivity measures are - mostly - recovered after estimating some form of a sales generating production function where output is proxied by sales. The standard approach has been to use the price index - of a given industry - to proxy for these unobserved prices. The use of the price index is only valid if all rms in the industry face the same output price and corresponds with the assumption that rms produce homogeneous products. In the case of dierentiated products this implies that the estimates of the input coecients are biased and in addition lead to productivity estimates that capture markups and demand shocks.2 In a second step these productivity estimates are then regressed on variables of interest, say the level of trade protection or taris. This implies that the impact on actual productivity cannot be identied - using a two-step procedure - which invalidates evaluation of the welfare implications. In this paper I analyze the productivity gains from trade liberalization in the Belgian textile industry. As in most empirical work that has addressed similar questions, I do not observe output at the plant-level and therefore unobserved prices and demand shocks need to be controlled for. In order to answer this question, I rst introduce a simple methodology for getting reliable estimates of productivity in an environment of imperfect competition in the product market where I allow for multi-product rms. The estimation strategy is related to the original work of Klette and Griliches (1996) where the bias of production function coecients due to using deated rm-level sales (based on an industry-wide producer price index) to proxy for rm-level output is discussed. In their application the interest lies in recovering reliable estimates for returns to scale and not in productivity estimates per se. At the same time a literature emerged allowing to correct for the simultaneity bias and recovering reliable estimates for productivity. The latter is a well documented problem when estimating a production function with OLS that inputs are likely to be correlated with unobserved productivity shocks and therefore lead to biased estimates of the production function. Olley and Pakes (1996) introduced an empirical strategy based on a theoretical dynamic optimization problem of the rm under uncertainty where essentially unobserved productivity in the production function is replaced with a polynomial in investment and capital. A series of papers used this approach to verify the productivity gains from changes in the operating environment of rms such as trade liberalization, trade protection among others. In almost all of the empirical applications the omitted price variable bias
Pavcnik (2002) documents the productivity gains from trade liberalization in Chile, Smarzynska (2004) nds positive spillovers from FDI in Lithuania and Van Biesebroeck (2005) nds learning by exporting in Sub-Saharan African manufacturing. Olley and Pakes (1996) analyze the productivity gains from deregulating the US telecom equipment industry. 2 Obtaining precise productivity estimates by ltering out price and demand shocks has a wide range of implications for other applied elds. For instance in applying recently developed methods to estimate dynamic (oligopoly) games where productivity is a key primitive (Collard-Wexler 2006).
1

was ignored or assumed away.3 In this paper I analyze productivity dynamics during a period of trade liberalization while correcting both for the omitted price variable and the simultaneity bias. I use the Olley and Pakes (1996) procedure to control for the simultaneity bias and their framework turns out to be very instructive to evaluate the importance of demand shocks in the production function and how they aect the productivity estimates.4 I empirically show that the traditional productivity measures still capture price and demand shocks which are likely to be correlated with the change in the operating environment. In addition, to correct for the omitted price variable bias and obtain unbiased coecients of the production function, I introduce a rich source of product-level data matched to the production dataset to analyze productivity dynamics in the Belgian textile industry during a signicant change in the trade regime by means of the abolishment of quota protection and increased quota levels. A large body of empirical work has studied the impact of various trade policy changes on productivity.5 By introducing a rich source of demand variation I am able to decompose the traditional measured productivity gains into real productivity gains and demand side related components. In order to evaluate whether opening up to trade is welfare enhancing, I make explicit assumptions on the nature of demand. This is in contrast to most papers dealing with the estimation of rm-level productivity where the assumptions on the nature of demand are often not mentioned but imply that all rms face the same price and produce homogenous products. In addition, the method sheds light on other parameters of interest - such as markups and the elasticity of substitution - and the role of dierences in product mix across rms. In the context of trade liberalization, a number of authors have found strong relationships between trade protection and markups. I nd that when correcting for unobserved prices and demand shocks, the productivity gains are more than halved and are small in magnitude. These results therefore shed light on the importance of both the productivity and markup response to a change in a trade regime. Some recent work has discussed the potential bias of ignoring demand shocks when estimating production functions based on deated rm-level sales to proxy for output. Katayama et al. (2003) start out from a nested logit demand structure and verify the impact of integrating a demand side on the interpretation of productivity. Melitz and Levinsohn (2002) assume a representative consumer with Dixit-Stiglitz preferences and they feed this through the Levinsohn and Petrin (2003) estimation algorithm.6 Foster, Haltiwanger and Syverson (2005) discuss the
3 Some authors did explicitly reinterpret the productivity measures as sales per input measures. For instance see footnote 3 on page 1264 of Olley and Pakes (1996). 4 This does not rule out the use of alternative proxy estimators such as the estimator suggested by Levinsohn and Petrin (2003), however, with some additional assumptions made on the relation between the unobserved productivity shocks and markups. See Appendix C for a discussion on this. 5 See Tybout (2000) for a review on the relationship between openness and productivity in developing countries. 6 The methodological part of this paper is closely related to Melitz and Levinsohn (2002). However, in addition to the plant-level dataset I have product-level information matched to the plants allowing me to put more structure on the demand side. They proxy the number of products per rm by the number of rms in an industry, while I observe the actual number of products produced by each rm and additional demand related variables. I use this additional source of variation to identify the elasticity of substitution. Aside from a discussion of the methodology,

relation between physical output, revenue and rm-level prices. They study this in the context of market selection and they state that productivity based upon physical quantities is negatively correlated with establishment-level prices while productivity based upon deated revenue is positively correlated with establishment-level prices. The few papers that explicitly analyze the demand side when estimating productivity or that come up with a strategy to do so all point in the same direction: estimated productivity still captures demand related shocks.7 The remainder of this paper is organized as follows. In section 2 the standard approach to estimate production functions is discussed. Furthermore, I introduce a demand system and show the bias on the production function coecients. Section 3 introduces the estimation strategy and the potential bias of using standard productivity estimates to evaluate policy changes. The readers interested in the empirical application can skip section 3 and are referred to sections 4-5. In section 4, I present the data that includes detailed product-level information in addition to a rich rm-level dataset of Belgian textile producers. In section 5 I present the coecients of the production functions as well as the estimated parameters of the demand system. In section 6 I analyze the eects of the trade liberalization episode in the EU textiles on productivity, where the trade liberalization is measured by the drastic fall in product specic quota protection. The quota information also serves as an important control variable for the unobserved prices through the introduction of the demand system. The last section concludes.

2
2.1

Estimating productivity using production data

Identication of the production function parameters

Let us start with the production side where a rm i at time t produces (a product) according to the following production function
q m k l Qit = L it Mit Kit exp(0 + it + uit )

(1)

where Qit stands for the quantity produced, Lit , Mit and Kit are the three inputs labor, materials and capital; and l , m and k are the coecients, respectively.8 The constant term 0 captures the mean productivity and captures the economies of scale, i.e. = (l + m + k ). Productivity is denoted by it and uq it is an i.i.d. component. The standard approach in identifying the production function coecients starts out with a production function as described in equation (1). The physical output Qit is then substituted f by deated revenue (R it ) using an industry price deator (PIt ). Taking logs of equation (1) and relating it to the (log of) observed revenue per rm rit = qit + pit , we get the following regression
I empirically show the bias in the production function coecients and in the resulting productivity estimates 7 See Bartelsman and Doms (2000) for a comprehensive review on recent productivity studies using micro data. Concerning the topic of this paper I refer to page 592. 8 The Cobb-Douglas production function assumes a substitution elasticity of 1 between the inputs. The remainder of the paper does not depend on this specic functional form. One can assume e.g. a translog production function and proceed as suggested below.

equation rit = xit + it + uq it + pit (2)

where xit = 0 + l lit + m mit + k kit . The next step is to use the industry wide price index pIt and subtract it from both sides to take care of the unobserved rm-level price pit . r eit = rit pIt = xit + it + (pit pIt ) + uq it (3)

Most of the literature on the estimation of productivity has worried about the correlation between the chosen inputs xit and the unobserved productivity shock it . The coecient on the freely chosen variables labor and material inputs will be biased upwards as a positive productivity shock leads to higher labor and material usage (E (xit it ) > 0). Even if this is corrected for, from equation (3) it is clear that if rms produce dierentiated products or have some pricing power the estimates of will be biased. As mentioned in Klette and Griliches (1996) inputs are likely to be correlated with the price a rm charges.9 The error term (uq it + pit pIt ) still captures rm-level price deviation from the average (price index) price used to deate the rm-level revenues. Essentially, any price variation (at the rm level) that is correlated with the inputs biases the coecients of interest () as E (xit (pit pIt )) 6= 0. A potential correlation is introduced when inputs are correlated with the rms output. The sign of the bias could go either way. Essentially the bias works through the correlation between the price a rm charges and its output and therefore introduces a correlation between the inputs and the unobserved price. Therefore rm-level inputs (materials and labor) are correlated with the unobserved price and thus under- or overestimates the coecients on labor and materials. This is referred to as the omitted price variable bias. Another source of bias is introduced by unobserved demand shocks that might lead to a higher price and induces a positive correlation between inputs and price. The omitted price bias might work in the opposite direction as the simultaneity bias - the correlation between the unobserved productivity shock ( it ) and the inputs (xit ) - making any prior on the total direction of the bias hard. It is also clear that even when the marginal product of the inputs () are not of interest, the productivity estimate is misleading as it still captures price and consequently demand shocks. The same kind of reasoning can be followed with respect to the measurement of material inputs where often a industry wide material price deator is used to deate rm-level cost of materials. However, controlling for unobserved prices takes - at least partly - care of this. The intuition is that if material prices are rm specic, a higher material price will be passed through a higher output price if output markets are imperfect, the extent of this pass through depends on the relevant markup. The only case where this reasoning might break down is when input markets are imperfect and output markets are perfectly competitive, which is not a very likely setup.10
The interpretation of the correlation is somewhat dierent here since my model is estimated in log levels and not in growth rates as in Klette and Griliches (1996). 10 If material prices dier across rms, an additional correlation of the input with the unobserved price pi is
9

2.2

Introducing demand and product dierentiation

I now introduce the demand system that rms face in the output market. The demand system is based on the standard horizontal product dierentiation model and allows for an unobserved quality component. The choice of demand system needed to identify the parameters of interest is somewhat limited due to missing demand data, i.e. prices and quantities. Therefore, one has to be willing to put somewhat more structure on the nature of demand. However, the modeling approach here does not restrict any demand system per se, as long as the inverse demand system can generate a (log-) linear relationship of prices and quantities. What follows then also holds for other demand systems and later on I will turn to some alternative specications. The inverse demand system is then used to substitute for the unobserved price variable in the revenue generating production function. I start out with single product rms and show how this leads to my augmented production function. In a second step I allow for rms to produce multiple products. The focus is on the resulting productivity estimates and in the case of multi-product rms these can be interpreted as average productivity across a rms products. 2.2.1 A Simple Demand Structure: Single Product Firms

I follow Klette and Griliches (1996) and later on I extend it by allowing rms to produce multiple products. I start out wit a simple (conditional) demand system where each rm i produces a single product and faces the following demand Qit Pit Qit = QIt exp(ud (4) it ) PIt where QIt is the industry output at time t, (Pit /PIt ) the relative price of rm i with respect to the average price in the industry, ud it is an idiosyncratic shock specic to rm i and is the substitution elasticity between the dierentiated products in the industry, where < < 1. The choice of this conditional demand system does not rule out other specications to be used in the remainder of the paper. However, it implies that the inverse of the elasticity of substitution (demand ) is the relevant markup as the substitution elasticity with respect to other goods (non textile products) is zero.11 The rms are assumed to operate in an industry characterized by horizontal product differentiation, where captures the substitution elasticity among the dierent products and is nite. As mentioned in Klette and Griliches (1996) similar demand systems have been used extensively under the label of Dixit-Stiglitz demand. The key feature is that monopolistic competition leads to price elasticities which are constant and independent of the number of varieties.
introduced through the correlation between output prices pi and material prices pm i . Note that this is in addition to the correlation between material mi used and prices pi .This follows from the fact that deated material costs m can be written as (mi + pm i pI ). 11 In the empirical analysis I replace the industry output QIt by a weighted average of the deated revenues, S i.e. QIt = ( i msit Rit /PIt ) where the weights are the market shares. This comes from the observation that a price index is essentially a weighted average of rm-level prices where weights are market shares (see Appendix A.2). Under the given demand structure it follows that (the rst order proxy for) the price index is a market share weighted average of the rm-level prices.

I refer to Berry (1994) for more on demand in industries with product dierentiation. It is clear that the demand system is quite restrictive and implies one single elasticity of substitution for all products within a product range and hence no dierences in cross price elasticities. In the empirical application the elasticity of substitution is allowed to dier among product segments. However, in most productivity studies the demand side is ignored and productivity is interpreted as an output per input measure (Katayama et al. 2003). The motivation for modeling demand explicitly here is to control for unobserved price variation. However the nal interest lies in an estimate of productivity and further relaxing the substitution patterns here would just reinforce the argument. Taking logs of equation (4) and writing the price as a function of the other variables results in the following expression where x = ln X pit = 1 (qit qIt ud it ) + pIt (5)

It is clear that if one does not take into account the degree of competition on the output market (rm price variation), that the analysis will be plagued by an omitted price variable bias and the estimated coecients are estimates of a reduced form combining the demand and supply side in one equation. I now further decompose the unobservable ud it in equation (4) into an unobserved demand shock ( it ) such as quality and an i.i.d. component to allow for unobserved rm (product) eects that impact demand. This leads to my general estimating equation of the revenue production function r eit = 0 + l lit + m mit + k kit + qIt + ( (7) it + it ) + uit
Exceptions are Dunne and Roberts (1992), Jaumandreu and Mairesse (2004), Eslava et al.(2004) and Foster et al. (2005) where plant-level prices are observed and thus demand and productivity shocks can be estimated separately. To my knowledge this is a very rare setup.
12

Now I only have to plug in the production technology as expressed in equation (1) and I have a revenue generating production function with both demand and supply variables and parameters. +1 +1 1 1 d (0 + l lit + m mit + k kit ) qIt + ( it + uq r eit = it ) uit

As discussed extensively in Klette and Griliches (1996) and Melitz and Levinsohn (2003), the typical rm-level dataset has no information on physical output per rm and prices.12 Commonly, we only observe revenue and we deate this using an industry-wide deator. The observed revenue rit is then substituted for the true output qit when estimating the production function. Ignoring the price thus leads to an omitted variable bias since it is not unlikely that a rms price is correlated with the used inputs. I now substitute expression (5) for the price pit in equation (2) to get an expression for revenue. From here forward, I consider deated revenue (r eit = rit pIt ) +1 1 1 qit qIt ud r eit = rit pIt = (6) it

 1 where h = (( + 1)/ )h with h = l, m, k ; = 1 , it = (( + 1)/ ) it , it = it and q 1 d uit = (( + 1)/ )uit uit . When estimating this equation (7) I recover the production function coecients (l , m , k ) and returns to scale parameter ( ) controlling for the omitted price variable and the simultaneity bias, as well as an estimate for the elasticity of substitution . In fact, to obtain the true production function coecients () I have to multiply the estimated reduced form parameters ( ) by the relevant markup ( +1 ). When correcting for the simultaneity bias I follow the Olley and Pakes (1996) procedure and replace the productivity shock it by a function in capital and investment. In my empirical analysis I will estimate various versions of (7) as the product information linked to every rm allows me to put more structure on the demand side, e.g. allowing the demand elasticity to vary across dierent segments and proxy for unobserved demand shocks ( it ) using product dummies. Bringing the extra information from the product space is not expected to change the estimated reduced form coecients ( ), but it will have an impact on the estimated demand parameter and hence on the true production function coecients ().13

2.2.2

Multi-product rms

I now allow rms to produce multiple products, where still captures the substitution elasticity among the varieties. For now I do not allow for dierent substitution patterns among products owned by a single rm as opposed to the substitution between products owned by dierent rms. The modelling approach here does allow for more realistic substitution patterns in the spirit of Berry, Levinsohn and Pakes (1995) among the various products produced and ultimately will be determined by the data at hand. The demand system is identical to the one expressed in equation (4), only a product subscript j is added. Note that the demand is now relevant at the product level. There are N rms and P M products in the industry with each rm producing Mi products, where M = i Mi .14 In the single product case the demand system is the same for every rm i, whereas in the multiple product case the demand is with respect to product j of rm i. Pijt s Qijt = Qst exp(ud (8) ijt + ijt ) Pst The demand for product j of rm i is given by Qij , QsIt is the demand shifter relevant at the
13 The setup is similar to the approach taken by Klette and Griliches (1996). However, three main problems remain unchallenged in their method, which are largely recognized by the authors. Firstly, industry output might proxy for other omitted variables relevant at the industry level such as industry wide productivity growth and factor utilization. The constant term and the residual in their model should take care of it since time dummies are no longer an option as they would take all the variation of the industry output. I use additional demand variables to control for demand shocks not picked up by industry output. Secondly, the residual still captures the unobserved productivity shock and biases the estimates on the inputs. I proxy for this unobserved productivity shock using the method suggested by Olley and Pakes (1996) to overcome the simultaneity bias, i.e. by introducing a polynomial in investment and capital. The third problem is closely related to the solution of the simultaneity problem. Klette and Griliches (1996) end up with a negative capital coecient partly due to estimating their production function in growth rates. 14 In the empirical application, I have 308 (N ) rm observations and 2,990 rm-product (M ) observations, with 563 unique product categories (j ).

product-level, Pst is the industry price index relevant at the product level, s is the demand elasticity relevant at the product space, ijt is unobserved demand shock at the product level 15 The elasticity of demand is (e.g. quality) and ud s ijt is product j specic idiosyncratic shock. now specic to a given product segment s of the industry. As mentioned above, the working assumption throughout this paper is that only the relevant variables at the rm level are observed, which is an aggregation of the product-level variables. This is the case in most of the studies using rm-level data to estimate a production function. However, as I will discuss later on in detail, I have information on the product market linked to the rm-level data which allows me to put somewhat more structure on the way the product-level demand and production are aggregated. Proceeding as in the single product case, the revenue of product j of rm i is rijt = pijt + qijt and using the demand system as expressed in equation (8) I get the following expression for the product-rm revenue rijt s + 1 1 s + 1 1 1 xijt qst + ( ijt + uq rijt pst = ijt ud (9) ijt ) s s s s s ijt I have assumed that the production function qij for every rm i for all its products Mi is given by the same production function (1) and it implies that the production technology for every product is the same and that no cost synergies are allowed on the production side. In Appendix B I relax this assumption and show a reduced form approach to allow for some spillovers in the production process. As before I substitute in the production technology as given by equation (1) where now a product subscript j is added. The aggregation from product to rm-level can be done in various ways and ultimately depends on the research question and the data at hand. If product specic inputs and revenues are available, the same procedure as in the single product rm applies, i.e. estimating a revenue production function by product j . However, observing revenue and output by product is hardly ever the case and so some assumptions have to be made in order to aggregate the product-level revenues to the rm level (the unit of observation in most empirical work). For simplicity I assume a constant demand elasticity across products () and I aggregate the product-rm revenue to the rm revenue by taking the sum over the number of products P i produced Mit , i.e. Rit = M j Rijt as in Melitz (2001). This leads to the following equation +1 1 r eit = 0 + l lit + m mit + k kit + qIt + np npit + it + + uit (10) | | it

where I have assumed that inputs per product are used in proportion to the number of products Xit (Xijt = M ) which introduces an additional term np npit where npit = ln(Mit ). The input it proportionality is driven by the lack of product-specic input data such as the number of employees that are used for a given product j . As mentioned above, in Appendix B I relax the
15

In the multi-product model I have to aggregate the revenues per product to the rms total revenue. The demand shifters are thus depending on the space, therefore I use the superscript s for the output and price index. In the empirical analysis - as in the single product case - I replace the output by the weighted average of deated segment revenues.

production aggregation from product to the rm level by essentially introducing a matrix that captures synergies from combining production of any 2 given segments within a single rm. Productivity and demand shocks are assumed to occur at the rm level and uit captures all the i.i.d. terms from both demand and supply (aggregated over products).16 The demand shifter qst is crucial as it allows me to identify the (segment specic) elasticity of demand through the assumption that it captures shocks in demand that are independent of the production function inputs and unobserved productivity. Furthermore it will turn out to be rm specic as I allow the demand elasticity to dier across products or segments of products. The latter is a result of allowing for rm specic product mixes and therefore each rm faces a (potential) dierent total demand over the various products it owns. In the case of a constant elasticity of demand across products (segments) and single product rms, this term is as before ( qIt ).

Estimation strategy and productivity estimates

I now briey discuss how to estimate the demand and production function parameters. Secondly, I allow for investment to depend on the unobserved demand shock ( it ) in the underlying Olley and Pakes (1996) model and I suggest a simple way (given the data I have) to control for this. Finally, I discuss the resulting productivity estimate and how it should be corrected for in the presence of product dierentiation and multi-product rms. I also provide a discussion on the importance of miss-measured productivity (growth) using the standard identication methods.

3.1

Estimation strategy: single and multi-product rms

Estimating the regression in (7) is similar to the Olley and Pakes (1996) correction for simultaneity, only now an extra term has to be identied.17 I group the two unobservables productivity it and demand shock it into one unobservable e it . Introducing the demand side clearly shows that any estimation of productivity also captures rm/product specic unobservables such as product quality for instance. I assume that the quality and productivity component follow the same stochastic process, i.e. a rst order Markov process with the same rate of persistence.18 Productivity is assumed to follow an exogenous process and cannot be changed by investment or other rm-level decision variables such as R&D or export behavior (De Loecker 2007).19 Both productivity and quality
Foster, Haltiwanger and Syverson (2005) do not observe inputs at the plant level, they observe product specic revenues which allows them to proceed by assuming that inputs are used in proportion given by the share of a given product in total rm revenue. 17 In the case of multi-product rms an additional parameter has to be identied. The identication depends on whether one allows the market structures to be dierent for single and multi-product rms. 18 A possible extension to this is to assume that the quality and the productivity shock follow a dierent Markov process. Therefore one can no longer collapse both variables into one state variable (see Petropoulus 2000 for explicit modeling of this). For now I assume a scalar unobservable (productivity/quality) that follows a rst order Markov process. However, I can allow for higher order Markov processes and relax the scalar unobservable assumption as suggested in Ackerberg and Pakes (2005), see later on. 19 Muendler (2004) allows productivity to change endogenously and suggests a way to estimate it. Buettner (2004) introduces R&D and models the impact of this controlled process on unobserved productivity. Acker16

10

are known to the rm when making its decision on the level of inputs. The new unobserved state variable in the Olley and Pakes (1996) framework is now e it = ( it + it ) and this is equivalent to Melitzs (2001) representation. Technically, the equilibrium investment function still has to be a monotonic function with respect to the productivity shock, e it , in order to allow for the inversion as in Olley and Pakes (1996) Here I have been more explicit on the nature of the unobservable e it containing both unobserved demand (quality) and productivity. However, it does not change the impact on investment. A rm draws a shock consisting of both productivity and quality and the exact source of the shock is not important as a rm is indierent between selling more given its inputs due to an increased productivity or increased quality perception of its product(s). We could even interpret investment in a broader sense, both as investment in capital stock and advertising. I replace the productivity e it component by a polynomial in capital and investment, recovering the estimate on capital in a second stage using non linear least squares. The demand parameters, labor and material are all estimated in a rst stage r eit = 0 + l lit + m mit + qIt + t (kit , iit ) + uit (12) it = it (kt , et) e t = ht (kt , it ) (11)

under the identifying assumption that the function in capital and investment proxies for the unobserved product/quality shock.20 A key parameter that I identify in this rst stage is the estimate of the markup ( ) which is identied by independent variation in demand shocks either at the industry (qIt ) or segment level (qst ) depending on the specication I consider. Note that the t (.) is a solution to a complicated dynamic programming problem and depends on all the primitives of the model like demand functions, the specication of sunk costs, form of conduct in the industry and others (Ackerberg, Benkard, Berry and Pakes; 2005). My methodology brings one of these primitives - demand - explicitly into the analysis and essentially adds explicitly information to the problem by introducing demand variables in the rst stage. Remember that this is required in order to recover estimates for true productivity ( it ) when rm-level prices are not observed. The identication of the capital coecient in a second stage will now improve as the estimate for it is now puried from demand shocks due to the introduction of demand variables in the rst stage. This is important as it is crucial to identify to the capital coecient. In a second stage (13) the variation in the variable inputs and the demand variation is subtracted from the deated revenue to identify the capital coecient. As in Olley and Pakes (1996) the news component in the productivity/quality process is assumed to be uncorrelated with capital in the same period since capital is predetermined by investments in the previous year. b kit ) + eit+1 r eit+1 bl lit+1 bm mit+1 b qIt+1 = c + k kit+1 + g ( it k (13)

berg and Pakes (2005) discuss more general extensions to the exogenous Markov assumption of the unobserved productivity shock. 20 Dynamic panel data econometrics uses lag structure and IV techniques to identify the production function parameters (Arellano and Bond, 1991).

11

where b is the estimate for out of the rst stage. Note that here I need to assume that unobserved demand and productivity shock follow the exact same Markov process in order to identify the capital coecient. If the demand shock does not follow the same process and is depending on productivity, identication is only possible through an explicit demand estimation as e.g. Berry, Levinsohn and Pakes (1995) in order to produce an estimate for it . Another way out is to assume that the unobserved demand shock is uncorrelated with capital and has no lag structure, but that would leave us back in the case where it is essentially ignored when estimating a revenue generating production function. Note that here I just spell out the identifying assumptions one always has to make when estimating a production function where revenue is used to proxy for output, regardless whether it is explicitly stated or not. The correction for the sample selection problem due to the non random exit of rms is as in the standard framework and leads to adding the predicted survival probability Pit+1 in g (.) in equation (13). The predicted probability is obtained from regressing a survival dummy on a polynomial in capital and investment. Productivity ( b it ) is then recovered by plugging in the estimated coecients in the produce tion function, (r eit bl lit bm mit bk kit b qIt ) e b it . +1 = The suggested framework does not rule out alternative proxies for the unobserved productivity shock. Levinsohn and Petrin (2003) use intermediate inputs as a proxy.21 Recently there has been some discussion of the validity of both proxy estimators. The rst stage of the estimation algorithm potentially suers from multicollinearity and the investment or material input function might not take out all the variation correlated with the inputs (Ackerberg, Caves and Frazer 2004). The criticism essentially comes from the assumptions of the underlying timing of the input decisions on labor and materials or investment. If indeed the rst stage would suer from multicollinearity, one can no longer invert the productivity shock and the estimates would not be estimated precisely.22 As noted by Olley and Pakes (1996), one can test whether the non parametric function used in the rst stage is well specied and is not collinear with labor by introducing the labor coecient in the last stage when identifying the capital coecient.23

3.2

Unobserved demand shocks and productivity

So far I have assumed that the unobservable e it - including both unobserved productivity and unobserved demand shocks (such as quality) - can be proxied by a non parametric function in investment and capital. The underlying assumption in that case is that investment proxies both the shocks in productivity ( it ) and unobserved demand shocks ( it ). I now relax this by allowing investment to explicitly depend on another unobservable - a demand shock - that varies across rms as suggested in Ackerberg and Pakes (2005). This notion also follows from
21 The choice among the dierent proxy estimators depends on many things such as the share of rms having non zero investments, and the assumptions one is willing to make. (Appendix C). 22 However, it is clear from the Ackerberg et al. (2004) that the OP procedure does not suer from this critique under the following timing assumptions: labor is chosen without perfect knowledge of the productivity shock. 23 Also see De Loecker (2007) where this test is implemented and the labor coecient is found to be insignicant e throughout all specications when running r hit +1 = c + k kit+1 + g (t k kit ) + c lit + eit+1 .

12

the discussion throughout the paper that both demand and production related shocks have an impact on observed revenue. Note that unobserved demand shocks would not enter the production function if we would observe physical output or rm-level prices when the investment policy function does not depend on say quality. However, when investment is allowed to depend on an unobserved demand shock (quality) as well, it enters through the productivity shock even when physical output or rm-level prices are observed. In this section we have a demand shock entering both through the investment policy function and through the use of revenue to proxy for output at the rm level. The details of the estimation thus depend on whether the demand shock (quality) enters both into the demand system and the investment function. In the empirical application I will estimate both versions using rmproduct dummies to control for unobserved product specic demand shocks. Note that in the case where the investment policy function does not depend on unobserved demand shocks, we recover parameters of product specic markups (see section 5.2.2). I refer to Appendix A.3 for a more detailed discussion on this.

3.3

Potential biases of using standard productivity estimates

When comparing with the standard approach to recover an estimate for productivity, it is clear that when estimating equation r eit = 0 + l lit + m mit + k kit + tf pit + uit (14)

where I denote tfpit as measured productivity, that the resulting productivity estimate (residual) is miss-measured. It captures demand shocks and product mix variation on top of the potentially dierently estimated coecients l , m , k and 0 . For now I assume away the unobserved demand shock it and focus on the unobserved productivity shock. The resulting measured productivity tfpit relates to the true unobserved productivity it in the following way (15) it = (tf pit qIt np npit ) +1 The estimated productivity shock consistent with the product dierentiated demand system and multi-product rms is obtained by substituting in the estimates for the true values ( , np and ). This shows that any estimation of productivity - including the recent literature correcting for the simultaneity bias (Olley and Pakes 1996 and Levinsohn and Petrin 2003) is biased in the presence of imperfect output markets and multi-product rms. Assuming an underlying product market a simple correction is suggested, i.e. subtract the demand variation and the number of products and correct for the degree of product dierentiation. One can even get the demand parameter out of a separate (and potentially more realistic) demand regression. Note that in the case of single product rms operating in a perfectly competitive market the estimated productivity corresponds to the true unobservable, given that the simultaneity and selection bias are addressed as well.

13

It is clear from equation (15) that the degree of product dierentiation (measured by ) only re-scales the productivity estimate. However, when the demand parameter is allowed to vary across product segments, the impact on productivity is not unambiguous. The number of products per rm Mi does change the cross sectional (across rms) variation in productivity and changes the ranking of rms and consequently the impact of changes in the operating environment or rm-level variables on productivity. In a more general framework of time varying number of products per rm (Mit ) the bias in measured productivity tfpit is given by (16). The traditional measure tf pit captures various eects in addition to the actual productivity shock it . t + 1 1 it + tf pit = t qIt + npt npit + (16) t | t | it Measured productivity consists of a pure demand specic term ( qIt ) and is related to the number of products, in addition to productivity and demand shocks interacted with the inverse of the markup and the Lerner index, respectively. This expression sheds somewhat more light on the discussion whether various competition and trade policies have had an impact on productive eciency. There is an extensive literature using a two stage approach where productivity is estimated in a rst stage and then regressed on a variable of interest. However, in the rst stage the relation of that variable of interest with demand related variance is omitted. Pavcnik (2002) showed that tari liberalization in Chile was associated with higher productivity, where essentially an interaction of time dummies and rm trade orientation was used to identify the trade liberalization eect on productivity.24 In terms of my framework, this measure of opening up to trade might also capture changes in the markup (through change in elasticity of demand) and the product mix of rms. Similar studies have essentially measured productivity in some form as expressed in equation (16). The increased (measured) productivity can be driven by four factors: i) increased product quality, ii) increased productivity, iii) more elastic demand and iv) increased number of products. Measuring increased productivity without taking into account the demand side of the output market and the degree of multi-product rms might thus have nothing to do with an actual productivity increase.25 Even in the case of single product rms measured productivity growth (tfpit ) captures demand shocks and changes in markups. Using equation (16) and assuming that rm i experienced no productivity gain at all ( it = it1 ) it is clear that we can still measure a productivity increase. These biased productivity (growth) measures are then regressed upon variables potentially capturing both cost and demand shifters making any conclusion drawn out of these set of regressions doubtful. It is straightforward to show the various biases one induces by using miss-measured productivity in a regression framework. Consider the following
24 I refer to this paper among a large body of empirical work as the analysis of productivity is done by controlling for the simultaneity bias and the selection bias as in Olley and Pakes (1996). 25 Harrison (1994) builds on the Hall (1988) methodology to verify the impact of trade reform on productivity and concludes that ... ignoring the impact of trade liberalization on competition leads to biased estimates in the relationship between trade reform and productivity growth .

14

regression equation where the interest lies in 1 verifying the impact of dit on measured tfpit tf pit = 0 + 1 dit + zit + it (17)

where zit captures a vector of control variables and it is an i.i.d. error term. Using expression (16) it is straightforward to verify the dierent sources of correlation that bias the estimate for 1 E (tf pit ) E ((qIt + it )/|t |) E (npit ) E ((t + 1)/ t ) it = + + (18) dit dit dit dit where the expectation is conditional upon zit . It is clear that impact of dit on productivity ( it ) is biased and the specic question and data at hand should help to sign the bias introduced by the various sources. For instance, if dit captures some form of trade liberalization (or protection), it is expected to have an impact on the industrys total output and elasticity of demand and results in a biased estimate for coecient 1 . Note that on top of the bias the point estimate of the productivity eect is multiplied by the inverse of the (rm specic) markup. Konings and Vandenbussche (2005) showed that markups increased signicantly during a period of trade protection after antidumping lings in various industries. The second term in (18) captures the correlation between the product mix and dit . Bernard, Redding and Schott (2003) suggest that an important margin along which rms may adjust to increased globalization and other changes in the competitive structure of markets is through changing their product mix. I will empirically verify the importance of this bias when evaluating the impact of decreased quota protection in the Belgian textile industry on estimated productivity in section 5.

The Belgian textile industry: Data and institutional details

I now turn to the dataset that I use to apply the methodology suggested above and in a later stage to analyze the trade liberalization process measured by a signicant drop in quota protection. My data capture the Belgian textile industry for the period 1994-2002. The rm-level data are made available by the National Bank of Belgium and are commercialized by BvD BELFIRST. The data contains the entire balance sheet of all Belgian rms that have to report to the tax authorities. In addition to traditional variables - such as revenue, value added, employment, various capital stock measures, investments, material inputs - the dataset also has information on entry and exit. FEBELTEX - the employers organization of the Belgian textile industry - reports very detailed product-level information on-line (www.febeltex.be). More precisely, they list Belgian rms (311) that produce a certain type of textile product. They split up the textile industry into 5 subsectors: i) interior textiles, ii) clothing textiles, iii) technical textiles, iv) textile nishing and v) spinning. Within each of these subsectors products are listed together with the name of the rm that produces it. From this source I was able to link rms with the number of dierent products they produce, including other information on the dierent segments of the textile industry. 15

I match the rms listing product information with the production dataset (BELFIRST) and I end up with 308 rms for which I observe both rm-level and product-level information.26 The average size of the rms in the matched dataset is somewhat higher than the full sample, since mostly bigger rms report the product-level data. Even though I loose some rms due to the matching of the product and the production datasets, I still cover 70 percent (for the year 2002) of total employment in the textile industry.27 From the BELFIRST dataset I have virtually the entire population of textile producers and this allows me to check for sample selection and sample representativeness. The entry and exit data are detailed in the sense that I know when a rm exits and whether it is a real economic exit, i.e. not a merger, acquisition or a break up into other rms. By adding the rich source of product-level data (FEBELTEX), it is clear that the industrial classication codes (NACE BELCODE) are sometimes incomplete as they do not necessarily map into markets. If one would merely look at rms producing in the NACE BELCODE 17, there would be some important segments of the industry left out, e.g. the subsector technical textiles also incorporates rms that produce machinery for textile production and these are not in the 17 listings. It is therefore important to take these other segments into the analysis in order to get a complete picture of the industry. Before I turn to the estimation I report some summary statistics of both the rm-level and product-level data. In Table 1 summary statistics of the variables used in the analysis are given. The average rm size is increasing over time (11 percent). In the last column the producer price index (PPI) is presented. It is interesting to note that since 1996 producer prices fell, only to recover in 2000. Sales have increased over the sample period, with a drop in 1999. However, measured in real terms this drop in total sales was even more sharp. Furthermore I also constructed unit prices at a more disaggregated level (3 digit NACEBELCODE) by dividing the production in value by the quantities produced and the drop in prices over the sample period is even more prevalent in specic subcategories of the textile industry and quite dierent across dierent subsectors (see Appendix A.). Together with the price decrease, the industry as whole experienced a downward trend in sales at the end of the nineties. The organization of employers, FEBELTEX, suggests two main reasons for the downward trend in sales. A rst reason is a mere decrease in production volume, but secondly the downward pressure on prices due to increased competition has played a very important role. This increased competition stems from both overcapacity in existing segments and
26 After matching the two sources of data it turns out that a very small fraction - 17 - of rms included in the FEBELTEX listing are also active in wholesale of specic textiles. I ran all specications excluding those rms since they potentially do not actually produce textile and all results are invariant to this. 27 A downside is that the product-level information (number of products produced, segments and which products) is time invariant and leaves me with a panel of rms active until the end of my sample period. Therefore I check whether my results are sensitive to this by considering a full unbalanced dataset where I control for the selection bias as well as suggested in Olley and Pakes (1996). The results turn out to be very similar as expected since the correction for the omitted price variable is essentially done in the rst stage of the estimation algorithm. The variation left in capital is not likely to be correlated with the demand variables and therefore I only nd slightly dierent estimates on the capital coecient.

16

from a higher import pressure from low wage countries, Turkey and China more specically.28 Export still plays an important role, accounting for more than 70% of the total industrys sales in 2002. A very large fraction of the exports are shipped to other EU member states and this is important as the quota restrictions are relevant at the EU level. The composition of exports has changed somewhat, export towards the EU-15 member states fell back mainly due to the strong position of the euro with respect to the British Pound and the increased competition from low wage countries. This trend has been almost completely oset by the increased export towards Central and Eastern Europe. The increased exports are not only due to an increased demand for textile in these countries, but also due to the lack of local production in the CEECs. To every rm present in this dataset I have matched product-level information. For each rm I know the number of products produced, which products and in which segment(s) the rm is active. There are ve segments: 1) Interior, 2) Clothing, 3) Technical Textiles, 4) Finishing and 5) Spinning and Preparing (see Appendix A. for more on the data). In total there are 563 dierent products, with 2,990 product-rm observations. On average a rm has about 9 products and 50 percent of the rms have 3 or fewer products. Furthermore, 75 percent of the rms are active in one single segment. This information is in itself unique and informative and ties up with a recent series of papers by Bernard et al. (2003) looking at the importance of dierences in product mix across rms. Table 2 presents a matrix where each cell denotes the percentage of rms that is active in both segments. For instance, 4.8 percent of the rms are active in both the Interior and Clothing segment. The high percentages in the head diagonal reect that most rms specialize in one segment, however rms active in the Technical and Finishing segment tend to be less specialized as they capture applying and supplying segments, respectively. This information is very interesting in itself, since it gives us some information about the product mix and product diversication. The last row in Table 2 gives the number of rms active in each segment. Again since rms are active in several segments, these numbers do not sum up to the number of rms in my sample. The same exercise can be done based on the number of products and as shown in Table 3 the concentration into one segment is even more pronounced. The number in each cell denotes the average (across rms) share of a rms products in a given segment in its total number of products. The table above has to interpreted in the following way: rms that are active in the Interior segment have (on average) 83.72 percent of all their products in the Interior segment. The analysis based on the product information reveals even more that rms concentrate their activity in one segment. However, it is also the case that rms that are active in the Spinning segment (on average) also have 27.2 percent of their products in the Technical textile segment. Firms active in any of the segments tend to have quite a large fraction of their products in Technical textiles, 8.27 to 27.7 percent. Finally the last two rows of Table 3 show the median and
An example is the ling of three anti-dumping and anti-subsidy cases against sheets import from India and Pakistan. Legal actions were also undertaken against illegal copying of products by Chinese producers (Annual Report of FEBELTEX; 2002). In section 5 I analyze the productivity dynamics during this increased competition period.
28

17

minimum number of products owned by a rm across the dierent segments. Firms producing only 2 (or less) products are present in all ve segments, but the median varies somewhat across segments (see Appendix A.1 for a more detailed description of the segments).

Estimated production and demand parameters

In this section I show how the estimated coecients of a revenue production function are reduced form parameters and that consequently the actual production function coecients and the resulting returns to scale change are underestimated. Furthermore, I allow for segment specic elasticity of substitution parameters and introduce product xed eects to further control for unobserved demand shocks.

5.1

The estimated coecients of augmented production function

I compare my results with a few base line specications: [1] a simple OLS estimation of equation (2), the Klette and Griliches (1996) specication in levels [2] and dierences [3], KG Level and KG Dif f respectively. Furthermore I compare my results with the Olley and Pakes (1996) estimation technique to correct for the simultaneity bias in specication [4]. In specication [5] I proxy the unobserved productivity shock by a polynomial in investment and capital and the omitted price variable is controlled for as suggested by Klette and Griliches (1996). Note that here I do not consider multi-product rms, I allow for this later when I assume segment specic demand elasticities. I replace the industry output QIt by a weighted average of the deated revenues, i.e. QIt = P ( i msit Rit /PIt ) where the weights are the market shares. This comes from the observation that a price index is essentially a weighted average of rm-level prices where weights are market shares (see Appendix A.2). Table 4 shows the results for these various specications. Going from specication [1] to [2] it is clear that the OLS produces reduced form parameters from a demand and a supply structure. As expected, the omitted price variable biases the estimates on the inputs downwards and hence underestimates the scale elasticity. Specication [3] takes care of unobserved heterogeneity by taking rst dierence as in the original Klette and Griliches (1996) paper and the coecient on capital goes to zero as expected (see section 1). In specication [4] we see the impact on the estimates of correcting for the simultaneity bias, i.e. the labor coecient is somewhat lower and the capital coecient is estimated higher as expected. The omitted price variable bias is not addressed in the Olley and Pakes (1996) framework as they are only interested in a sales per input productivity measure. Both biases are addressed in specication [5] and the eect on the estimated coecients is clear. The correction for the simultaneity and omitted price variable go in opposite direction and therefore making it hard to sign the total bias a priori. The estimate on the capital coecient does not change much when introducing the demand shifter as expected since the capital stock at t is predetermined by investments at t 1, however, it is considerably bigger than in the Klette and Griliches (1996) approach. The correct estimate 18

of the scale elasticity (l +m +k ) is of most concern in the latter and indeed when correcting for the demand variation, the estimated scale elasticity goes from 0.9477 in the OLS specication to 1.1709 in the KG specication. The latter specication does not take control for the simultaneity bias which results in upward bias estimates on the freely chosen variables labor and material. This is exactly what I nd in specication [5], i.e. the implied coecients on labor drops when correcting for the simultaneity bias (labor from 0.3338 to 0.3075).29 The estimated coecient on the industry output variable is highly signicant in all specications and is a direct estimate of the Lerner index. I also show the implied elasticity of demand and markup. Moving across the various specications, the estimate of the average Lerner index (or the markup) increases as I control for unobserved rm productivity shocks. Moving from specication [2] to [3] I implicitly assume a time invariant productivity shock which results in a higher estimated Lerner index (from 0.2185 to 0.2658). In specication [5] productivity is modelled as a Markov process and no longer assumed to be xed over time. Controlling for the unobserved productivity shock leads to a higher estimate of the Lerner index (around 0.30) as the industry output variable no longer picks up productivity shocks common to all rms as proxied by investment and capital. Finally, an interesting by-product of correcting for the omitted price variable is that I recover an estimate for the elasticity of demand and for the markup. The implied demand elasticities are around 3 and the estimated markup is around 1.4.30 These implied estimates are worth discussing for several reasons. First of all, this provides us with a a check on the economic relevance of the demand model I assumed. Secondly, the implicit working assumption in most empirical work is that = and the estimates here provide a direct test of this. Thirdly, they can be compared to other methods (Hall 1988 and Roeger 1995) that estimate markups from rm-level production data. The message to take out of this table is that both the omitted price variable and the simultaneity bias are important to correct for, although that the latter bias is somewhat smaller in my sample. It is clear that this will have an impact on estimated productivity. The estimated reduced form parameters ( ) do not change much when controlling for the omitted price variable in addition to the simultaneity bias correction since the control is (in these specications) not rm specic. However, it has a big impact on the estimated production function parameters (), which by itself is important if one is interested in obtaining the correct marginal product of e.g.
Note that here my panel is only restricted to having rms with observations up to the year 2002 in order to use the product-level information and thus allows for entry within the sample period. However, as mentioned before my estimates of the production function are robust to including the full sample of rms. To verify this, I estimate a simple OLS production function on an unbalanced dataset capturing the entire textile sector. The capital coecient obtained in this way is 0.0956 and is very close to my estimate in the matched panel (0.0879), suggesting that the sample of matched rms is not a particular set of rms. 30 Konings, Van Cayseele and Warzynski (2001) use the Hall (1988) method and nd a Lerner index of 0.26 for the Belgian textile industry, which is well within in the range of my estimates (around 0.30). They have to rely on valid instruments to control the for the unobserved productivity shock. A potential solution to overcome this is a method proposed by Roeger (1995) were essentially the dual problem of Hall (1988) is considered to overcome the problem of the unobserved productivity shock, however one is no longer able to recover an estimate for productivity.
29

19

labor. The industry output variable captures variation over time of total deated revenue and as Klette and Griliches (1996) mentioned therefore potentially picks up industry productivity growth and changes in factor utilization. If all rms had a shift upwards in their production frontier, the industry output would pick up this eect and attribute it to a shift in demand and lead to an overestimation of the scale elasticity. In my approach, the correction for the unobserved productivity shock should take care of the unobserved industry productivity growth if there is a common component in the rm specic productivity shocks ( i ). In the next section I use the product-level information that allows for rm specic demand shifters as rms have dierent product portfolios over the various segments of the industry. Potentially, estimated productivity is dierent due to dierent estimated parameters ( ) and additional terms (industry output) capturing the shifts in demand for the products of a rm in a given segment of the industry. The estimated coecients on the inputs ( ) are not expected to change much once more demand information is introduced since they capture both demand and production variation, the implied production function parameters () however are expected to change.

5.2

Segment specic demand, unobserved product characteristics and pricing strategy

So far, I have assumed that the demand of all the products (and rms) in the textile industry face the same demand elasticity and I have assumed that the demand shock ud ijt was a pure i.i.d. shock. Before I turn to the productivity estimates, I allow for this elasticity to vary across segments and I introduce product dummies. In Appendix A.2 I present the evolution of producer prices in the various subsectors of the textile industry and it is clear that the price evolution is quite dierent across the subsectors suggesting that demand conditions were very dierent across subsectors and from now on I consider the demand at the segment level. Firstly, I construct a segment specic demand shifter - segment output deated - and discuss the resulting demand parameters. Secondly, I introduce product dummies to control for product specic shocks, essentially proxying for j . Finally, I split up my sample according to rms being active in only 1 or more segments. Firms producing in several segments can be expected to have a dierent pricing strategy since they have to take into account whether their products are complements or substitutes. Note that here the level of analysis is that of a segment, whereas the pricing strategy is made at the individual product level. 5.2.1 Segment specic demand parameters

The demand parameter is freed up to be segment-specic by interacting the segment demand shifter (segment output) with the segment share variables.31 The share variable Sis is the fraction
I have also estimated demand parameters one level deeper, see Appendix A.1 for the structure of the segments. This leads to a model with 51 dierent demand elasticities and identication is somewhat harder as the number of observations for some of the products is insucient. However, for a set of subsegments I recover signicant and meaningful estimates for markups.
31

20

P of rm i s products in segment s (Mis ) in the rms total number of products (Mi = s Mis ), where s = {1 (Interior), 2 (Clothing), 3 (Technical ), 4 (Finishing), 5 (Spinning and Preparing) }. Note that the demand elasticity is now identied using rm specic variation as the share variable is rm specic. As was shown in Tables 3 and 4, using the product information revealed a pattern of activity concentration into one segment on average, however there is quite some variation across rms.32 I now turn back to the general setup of the paper with multi-product rms. The demand for every product is given by (8) and qst captures the product specic demand shifter. As in the single product rm case I proxy the demand shifter by output, however, now it is segment output. The segment output I consider is constructed in the following way. I observe rm-level revenue rit and I know the share of the rms products per segment in its total products produced (Sis ). I consider the revenue of rm i in segment s to be Rist = Rit Sis with Sit = Mis /Mi . That is, if a rm has 20 percent of its products in segment 1 (Interior Textiles) I assume that 20 percent of its revenue comes from that segment. The relevant weight to construct the segment output is Rist vist = SN , where Ns is the number of rms active in segment s. The segment output qst is s i Rist P eist as before. I now introduce these terms interacted with the segment then proxied by i vist r share variable in the augmented production function and estimate the following regression ! 5 X r eit = 0 + l lit + m mit + k kit + s qst Sis + np npit + (19) it + uit
s=1

I present the estimated coecients s and the distribution of the estimated demand parameter in Table 5. One can immediately read of the implied demand parameters for the various segments in the textile industry for those rms having all their products in one segment (Sis = 1). Introducing multi-product rms in this framework explicitly implies a correction for the number of products produced. As mentioned before, since I do not observe the product specic inputs at the rm level, I have assumed that the product specic input levels are proportional to the total rm input, where the proportion is given by the number of products produced (ln Mi = npi ). The coecient on this extra term is negative and highly signicant of 0.0396.33 The rst row in Table 5 shows the estimated coecients implying signicantly dierent demand parameters for the various segments. I also include the implied demand parameters relevant for rms having all their products in a given segment. For instance, rms having all their products in the segment Interior face a demand elasticity of 5.2966. In panel B of table 5 I use the rm specic information on the relative concentration (Sis ) and this results in a rm specic elasticity of demand and markup which are in fact weighted averages over the relevant segment parameters. I stress that this comes from the fact that rms have multiple products across dierent segments and therefore the relevant demand condition is dierent for every rm.34
As mentioned before, I do not observe the change of the product mix over time. It is reassuring, however, that based on the US Census data (Bernard et al. 2003) rms only add or drop about 1 product over a ve year period, or less than 2 products over a nine year period which corresponds to my sample length (1994-2002). To the extend that this variation is not picked up by the proxy for i , it potentially biases the input coecients. 33 I also used the number of segments and the results are similar. 34 The same is true for the estimated production function coecients, since they are obtained by correcting for
32

21

5.2.2

Unobserved product characteristics

I now introduce product dummies to control for product specic unobserved demand shocks ( j ). In terms of section 3.2 the product dummies proxy for the unobserved demand shock quality - that is rm specic and potentially impacts the investment decision. I assume time invariant unobserved product characteristics. As mentioned above, there are 563 products (K ) in total (and a rm produces 9 of these on average) which serve as additional controls in the rst stage regression (20). The product dummies are captured by P RODik where P RODik is a dummy variable being 1 if rm i has product k. Note that I introduce the product dummies motivating the need to correct for product specic demand shocks such as unobserved quality. However, they will also capture variation related to the production side and those two types of variations are not separable.35 The identifying assumption for recovering an estimate on the capital coecient is that productivity and the unobserved demand shock are independent. However, using the product dummies in the proxy for productivity, the identifying assumption becomes less strong, i.e. I lter out time invariant product unobservables. Note that in the standard approach for identifying the production coecients, demand variation is not ltered out, both observed and unobserved. Here I allow for product unobservables and demand shocks to impact investment decisions, on top of proxying for the demand shocks proxied by segment output and product dummies. r eit = 0 + l lit + m mit +
5 X s=1

In Table 5 I show that the demand parameters do not change too much as expected, as well as the production related coecients. However, the point estimates are more precise and 62 out of the 652 products are estimated signicantly dierent from the reference product conrming the importance of controlling for time invariant product characteristics. As mentioned above the interpretation of these coecients is somewhat harder as the product dummies are introduced to proxy for unobserved demand shocks, however, they will also pick up product-specic production related dierences. As stressed before, all these extra controls come into play if the interest lies in getting an estimate on productivity taking out demand related variation. In terms of economic interpretation, Table 5 suggests that rms operating in the Finishing segment (only) face less elastic demand. The high elastic demand segments are Interior and Spinning capturing products - like linen, yarns, wool and cotton - facing high competition from low wage countries.36 In Appendix A.1 I relate these demand parameters to changes in output
the degree of production dierentiation which is rm specic ( i ). 35 I introduce the product dummies without interactions with the polynomial terms in investment and capital since that would blow up the number of estimated coecients by K . This then coincides with assuming that the quality unobservable does not enter the investment policy function in the rst stage and just correcting for the demand unobservable. However, it matters for the second stage, i.e. this variation is now not subtracted from deated sales (r h) like the variable inputs. This would imply that the time invariant product dummies would proxy the unobserved demand shock completely. Therefore, the resulting productivity will still capture a time variant demand shocks - say improved product quality - component. 36 Increased international competition in the Interior and Spinning segments is documented in section 6 where

et (iit , kit , P RODi1 , ..., P RODiK ) + uit (20) s qst Sis + np npi +

22

prices at more disaggregated level and I nd that indeed in those sectors with relative high elastic demand, output prices have fallen considerably over the sample period. 5.2.3 Single versus multi-product rms

So far I have assumed that the pricing strategy of rms is the same whether it produces one or more products, or whether it is active in one or more segments. Remember that the revenue observed at the rm-level is the sum over the dierent product revenues. Firms that have products in dierent segments are expected to set prices dierently since they have to take into account the degree of complementarity between the dierent goods produced. I relax this by simply splitting my sample according to the number of segments a rm is active in. The underlying model of price setting and markups can be seen as a special case where own and cross elasticities of demand are restricted to be the same within a segment. In the third row of Table 5 I present the estimated demand parameters for rms active in only 1 segment and for those active in at least 2. As expected the estimated demand elasticities for the entire sample are in between both. Firms producing products in dierent segments face a more elastic (total) demand since a price increase of one of their product also impacts the demand for their other products in other segments.37 This is not the case for rms producing only in 1 segment, leading to lower estimated demand elasticities. It is clear that the modeling approach here does allow for various price setting strategies and dierent demand structures. From the above it is clear that productivity estimates are biased in the presence of imperfect competitive markets and ignoring the underlying product space when considering rm-level variables. It is clear that the data and the research question at hand will dictate the importance of the various components captured by traditional productivity estimates. In the next section I analyze the productivity gains from the trade liberalization in the Belgian textile industry and I compare my results with the standard productivity estimates, which are in fact sales per input measures and not necessarily lead to the same conclusions.

Trade liberalization and productivity gains

In this section I introduce product-level quota restrictions as additional controls for the unobserved rm-level price variable in the demand system and consequently in the augmented production function. In section 5 I showed that the industry output variable was highly signicant, however, it implied rather high returns to scale estimates. Including the quota restriction variable is expected to lead to somewhat lower estimates on the industry output variable QI . Firms protected by quota are expected to have higher market share - if anything - and produce more. I will correct for the potential upward bias in the Lerner index. In addition the quota
quota protection is discussed. 37 Note that now the implied demand elasticities are given by the weighted sum over the various segments a rm is active in, where weights are the fraction of the number of products in a segment in the total number of products owned.

23

variable will control for additional variation in unobserved rm-level prices as producers are expected to be able to set higher prices if import is restricted even more so since quota tend to apply on suppliers with lower costs of production (wages). First I introduce the quota data and discuss how it relates to the rm-level data. Secondly, I introduce the quota restriction measure into the augmented production function. The resulting estimated productivity is then used to verify to what extent that abolishing the quota on imports has contributed to within-rm productivity gains in the Belgian textile industry and how results using standard techniques to estimate productivity dier from the methodology suggested in this paper. In contrast to within-rm productivity changes, aggregate industry productivity can increase by the mere exit of lower productivity rms and/or the reallocation of market share towards more productive rms.38 As shown in Syverson (2004), demand shocks might in turn impact the aggregate productivity distribution.

6.1

The quota data: raw patterns and a measure for trade liberalization

The quota data comes straight from the SIGL database constructed by the European Commission (2003) and is publicly available on-line (http://sigl.cec.eu.int/). Note that this data is at the EU level since Belgium has no national wide trade policy and so quota at the EU level are the relevant quota faced by Belgian producers. This database covers the period 1993-2003 and reports all products holding a quota. For each product the following data is available: the supplying country, product, year, quota level, working level, licensed quantity and quantity actually used by the supplying country.39 From this I constructed a database listing product-country-year specic information on quota relevant for the EU market. Before I turn to the construction of a variable capturing the quota restriction relevant at the rm level, I present the raw quota data as it shows the drastic changes that occurred in trade protection during my sample period 1994-2002. In addition to observing whether a given product is protected by a quota, the level of allowed import quantities measured in kilograms (kg) or number of pieces - depending on the product category - is provided. In total there are 182 product categories and 56 supplying countries, where at least one quota on a product from a supplier country in a given year applies. In terms of constructing a trade liberalization or protection measure various dimensions have to be considered. A rst and most straightforward measure is a dummy variable that is 1 if a quota protection applies for a certain product category g on imports from country e in year t (qregt ) and switches to zero when the quota no longer applies. However, increasing the quota levels is also consistent with opening up to trade and thus both dimensions are important to look at. Table 6 below shows the number of quota that apply for the sample period 1994-2002. In addition I provide the average quota levels split up
38 It is clear that decompositions of aggregate industry productivity using biased measures of rm-level productivity will provide dierent answers as to how important net entry, reallocation and within productivity growth are. In fact given the framework suggested here, it is easy to show how we over- (under) estimate the various components of aggregate productivity. Under the empirical relevant scenario that entrants charge lower prices, it is clear that the importance of entry is underestimated since tf pit = it + (pit pIt ). 39 Appendix A.4 describes the quota data in more detail and provides two cases on how quota protection changed.

24

in kilograms and number of pieces, both expressed in millions. It is clear from the second column that the number of quota restrictions have decreased dramatically over the sample period. By 2002 the number of quota fell by 54 percent over a nine year period and these numbers refer to the number of product-supplier restricted imports. Columns 3 to 6 present the evolution by unit of measurement and the same evolution emerges: the average quota level increased with 72 and 44 percent for products measured in kilograms and number of pieces, respectively. Both the enormous drop in the number of quota and the increase in the quota levels of existing quota, points to a period of signicant trade liberalization in EU textile industry. It is essentially this additional source of demand variation I will use to identify the demand parameters in the augmented production function and verify how this gradual opening up to foreign textile products has impacted rm-level productivity. As mentioned above the product classications in the quota data are dierent from the rmlevel activity information and have to be aggregated to the rm-level revenue and input data. The average quota restriction (qr) that applies for a given product g is given by qrgt = X
e

aet qregt

(21)

where aet is the weight of supplier e in period t. This measure is zero if no single quota applies to imports of product g from any of the supplying countries at a given time, and one if it holds for all supplying countries. A nal step is to relate the quota restriction measure to the information of the rm revenue and production data. The 182 dierent quota product categories map into 390 dierent 8 digit product codes. The latter correspond to 23 (l) dierent 4 digit industry classications (equivalent to the 5 digit SIC level in the US) allowing me to relate the quota restriction variable to the rm-level variables. Aggregating over the dierent product categories leaves me with a quota measure of a given 4 digit industry code. I consider the average across products within an industry l (qrit ) where a rm i is active in as given by (22) and the number of quota (nqrit ) that apply in a given industry l. qrit = 1 X qrgt Ngt
g l(i)

(22)

In Figure 1 I show the evolution of the quota restriction variable given by (22) split up by segment. Again the same picture emerges, in all segments the average quota restriction has gone down considerably over the sample period, however, there are some dierences across the various segments and it is this variation that will help to estimate the segment specic demand elasticities. The construction of the quota restriction measures provides me with an additional control for the unobserved price variable and it is assumed to enter the demand system (4) as a residual demand shock ud it . This implies that it is also assumed to be independent (conditional on the other controls) with any of the production data variables.

25

6.2

Introducing quota restrictions in the demand system

I now introduce the average quota restriction qrit and the number of quota nqrit into the unobserved part of the demand system. The interpretation of this model is to estimate the elasticity of substitution (demand) that is consistent with international competition. It implies that the intercept for each rm is allowed to dier according to the protection of its products.40 I allow for segment specic demand elasticities and multi-product rms and I control for time invariant unobserved product eects using product dummies. This leads to the following augmented production function (23) r eit = 0 + l lit + m mit + +t (iit , kit ) +
K X k=1 5 X s=1

s qst Sis + ( qr qrit + nq nqrit ) + np npi (23)

k P RODik + uit

where the term in parentheses captures the quota measures.41 Before I present the estimated coecients, I note that introducing the quota restriction information helps estimating the s and potentially the production function parameters. Table 7 presents the estimated Lerner indices ( ) and compares them with the specication where the extra demand variation captured by the quota restriction ( qr qrit + nq nqrit ) is not included. I also recover product specic estimates and about 20 products are estimated signicantly dierent from their respective segment average (see Appendix A.3). These can be interpreted as product specic Lerner estimates under the assumption that the investment decision does not depend on the unobserved demand shock. The last 4 rows show the estimated production coecients and the implied returns to scale. The estimate on the quota restriction variable immediately provides information on how standard estimated productivity estimates incorporate demand shifters.42 As expected, the coecients on the segment output are estimated lower conrming the prior that the quota restriction measure is positively correlated with the segment output, i.e. higher protection, higher domestic production. As noted by Tybout (2000), the eect from restricting imports is that rms might exploit their enhanced market power and that protection is likely to increase the market size for domestic producers.
I have also estimate a change in the slope of the demand curve (elasticity). The identication is somewhat harder as rms can be active in dierent segments experiencing dierent changes in the protection, however, the results are invariant. 41 A well documented problem of using trade liberalization or protection measures in a regression framework is that they are potentially endogenous as rms might lobby for protection. In order to verify whether in my sample producers of certain product categories were able to keep higher level of protection, I ran a regression of qrge2003 on qrge1993 (N = 1, 097) nding a strong negative relation which suggests that protection in all product categories decreased over time. In addition, when analyzing the productivity eects I include product category (l) dummies controlling for (time invariant) dierences in lobbying-for-protection activities across producers active in dierent product categories. 42 All the results are based on unweighted averages. I have also experimented using the quota levels to construct the weights. These would capture the importance of a given quota protection in the overall import restriction and the extent to which import demand for a given product can be substituted away to another supplier. Due to the dierent unit of measurement in the levels, the interpretation of a change in qr is less clear.
40

26

The estimates on the inputs are very similar after introducing the additional demand information as expected, since these are just reduced form parameters. However the implied production coecients do change since the estimated demand elasticities change and this is reected in the lower estimated returns to scale. Note that the capital coecient is estimated lower compared to Table 4 where no product-rm dummies were used. The latter capture time invariant product dierences and improves the estimation of the capital coecient by purifying the error term in the nal stage (13) from any product-rm specic time invariant unobservables capturing quality dierences on top of the observed demand variation across segments. Finally, in Appendix A.3 I verify whether the estimates of the segment specic markups are sensitive to the underlying assumptions of the production process such as the timing of inputs with respect to the productivity shock and the substitution elasticity among inputs and I nd that my estimated demand parameters are robust to this.

6.3

The impact of relaxing trade protection on productivity

The coecient on the quota restriction variable is estimated highly signicant and with a negative sign, -0.0886. As previous studies have shown productivity gains are associated with trade liberalization, although measured in dierent ways essentially establishing a highly signicant positive correlation between productivity and opening up to trade.43 The estimated productivity shock in a standard OP setup would then still include markups and demand shifts introduced by the change in trade policy. Therefore it is crucial to purify productivity estimates from the price and demand related variation in order to get at the true impact of trade liberalization on productivity and productivity growth. The distinction between both is important as to know whether opening up to trade does impact productivity growth and hence has a long run impact on the eciency of an economy. The interpretation in my specication is somewhat more complicated. To the extent that the polynomial in capital and investment picks up the unobserved productivity shock, the quota restriction variable picks up demand shocks. However, it is clear that it will also pick up variation related to true productivity that is not controlled for by the polynomial in investment and capital. It is exactly the relation between productivity and the trade liberalization measure that is of interest. In order to verify the extent to which trade liberalization - measured by a decrease in quota restrictions - has impacted the productivity of Belgian textile producers I follow the standard 2 stage approach and show how the results change when using my corrected productivity estimates. I consider the following regression b it = 0 + 1 qrit + 2 nqrit + it (24)

where b refers to the estimated productivity and I will consider various versions of (24). In all regressions I include quota product classication dummies (23 categories) capturing time
43

See Tybout (2000) for an overview and e.g. Pavcnik (2002) for a country study.

27

invariant productivity (growth) dierences among categories. Table 8 presents the estimates of 1 under various specications. Before I turn to each specication, it is clear that - across all specications - using the standard OP productivity estimate leads to an overestimation of the impact of trade liberalization.44 Note that a decrease in the quota restriction variable corresponds with less quota protection or opening up of trade. So a negative coecient implies productivity gains from relaxing quota restrictions. In all specications the sign is negative and highly signicant and the interpretation of the coecient is the productivity gain for abolishing quota on all products from all countries. Specication I is the level regression and implies a 6.37 percent higher productivity for rms not protected by a single quota and using OP the estimate is much higher, 10.68 percent. Given the Markov assumption of productivity in the estimation algorithm and knowing that rm productivity estimates are highly persistent over time, specication II introduces lagged productivity as a regressor. The impact of the quota restriction variable is estimated more precise and somewhat lower. In specication III and IV , I run the regression in growth rates revealing the same pattern as in specication I . In specication IV , however, I include lagged levels of the quota restriction variable. Controlling for the lagged levels of the quota restriction measure leads to a higher point estimate on 1 , showing that the impact of relaxing quota restrictions on productivity depends on the initial level of the quota. If the quota protection was initially low, there is not much impact on productivity. Specication V considers long dierences (3 year period) and the results are robust to this, although estimated somewhat less precise due to the signicant drop in observations. In order to recover an estimate of the elasticity of productivity with respect to quota restrictions, I evaluate this at the mean (of the change in quota restriction) by segment. Table 9 shows the impact of a 10 percent decrease in the quota restriction measure on productivity for the various segments and compares it with the results using the OP productivity estimates. A 10 percent decrease in the quota restriction measure can come about by products being no longer protected from all or some supplying countries.45 As established in the previous table trade liberalization leads to higher productivity, however, there are some dierences across segments. A 10 percent decrease in my quota restriction measure leads only to a 1.6 percent higher productivity in the Finishing segment, as opposed to a 4.37 percent increase in the Interior segment. This result is what one would expect given the dierent paths of the quota restriction variable by segment as shown in Figure 1. The Finishing segment started out with a relatively low level of protection in 1994 (0.3) and stays rather at after 1996. The other segments - with higher estimated elasticities - had much higher levels of protection initially, e.g. the Interior segment was highly protected (qr = 0.85) in 1994 and by 2002 protection was signicantly lower (qr = 0.3). It is clear that the productivity gains
Using the estimates one can derive that the segments with a relative high level of protection have higher markups as expected (e.g. Tybout 2000). This - together with the scaled point estimate - leads to a biased estimate of the eect of relaxing quota protection on standard estimated productivity (OP). 45 The average quota restriction measure is 0.43 and the average change in this measure is -0.05, which is around 10 percent.
44

28

are much smaller (more than halved) and this is what one would expect for rms operating in an advanced economy, as opposed to rms active in more developing regions. The results show that decomposing the residual from a sales generating production function into productivity and demand related factors, is important to evaluate the impact of trade liberalization on productivity. Furthermore in Table 9, I present the elasticities evaluated at the mean of the change in the quota restriction for two dierent periods, 1994-1997 and 1998-2002. The rst period is characterized by a sharper fall in the quota protection (see Table 9) and therefore leads to higher estimated elasticities. The sharp fall of the number of quota in the period 1994-1997 is consistent with the process of the preparation of EU enlargement towards Central and Eastern Europe (CEE). By the year 1998 almost all trade barriers between the EU and the candidate countries of CEE were abolished as part of the Europe Agreements (EC 2005). The Europe Agreements were setup to establish free trade in industrial products over a gradual, transition period. This implied that industrial products from the associated countries (mostly CEE) have had virtually free access to the EU since the beginning of 1995 with restrictions in only a few sectors, such as agriculture and textiles. However, even in the last period (1998-2002) the productivity gains are still estimated around 3 percent with the exception of the Finishing segment which had a relatively low level of quota protection throughout the sample period. Finally, as mentioned before another measure of relaxing quota restrictions is by increasing the level of existing quota. In order to verify the impact of this on productivity I consider only those industry categories (4 digit NACE) that have some form of protection, i.e. where I observe a positive level of protection and the unit of measurement of a quota level is constant within a given industry code (23 categories). This dimension of opening up to trade has been the predominant strategy for the EU when it comes to imports from outside CEE and other new EU member states and not as much through abolishing quota. In Table 10 I list the supplying countries where relaxing import restrictions mainly occurred through higher levels of quota. I report the increase in the average level per quota during my sample period 1994-2002. The countries listed have gained access to the EU textiles market under a signicant increase of quota levels. For instance the average quota level on textile products from Pakistan has more than doubled over a nine year period (129 and 144 percent depending on product category). This process is not captured by the quota restriction variable qr that picks up whenever a quota on a given product from a supplying country is abolished. In order to verify the impact of increased quota levels - in addition to the abolishment of quota - I include a variable that measures the total level of quota (in logs) in a given industry in the regression framework of specication II . Specication V I in Table 10 shows the results of including the level variable. The quota restriction variable has a negative sign as before and the coecient on the level variable is estimated with a positive sign: an increase in the level of quota is consistent with increased competition from foreign textiles products and has a positive eect on productivity. The point estimate is an elasticity and implies that if quota levels increase by 29

10 percent that productivity increases with 1.9 percent. The simultaneous abolishment of quota protection and the increase in the quota levels are associated with higher productivity of Belgian textile producers. Productivity gains were higher for rms active in segments which initially were highly protected as they had to restructure more in order to face the increased competition from non-EU textile producers. However, the magnitude of the productivity gains are fairly small compared to those obtained with standard techniques. As mentioned before, the results presented in Table 10 can be interpreted as a decomposition of measured productivity gains from relaxing trade protection into true productivity gains and demand shocks. Here, I nd that around 50 percent is only picking up actual productivity gains.

Conclusion

In this paper I suggest a method to correct for the omitted price variable in the estimation of productivity. I have introduced a simple demand side and I explicitly allow rms to have multiple products. I introduce a simple aggregation from product space into rm space and derive a straightforward estimation strategy. I show that measured productivity increases need not to reect actual productivity increase. This casts some doubt on the recent empirical ndings that link changes in the operating environment - such as trade protection - on rmlevel productivity (growth) in a two-stage approach. I illustrate this methodology by analyzing productivity in the Belgian textile industry using an unique dataset that in addition to rm-level data has product-level information. Adding extra product-level information to the plant-level data appears to be a successful rst step in separating out demand variation and product mix from estimated productivity. The results here are obtained using a tractable and fairly standard demand system. The extent to which the results established in this paper are robust to using a richer demand system is ultimately an empirical question. However, it is clear that independent of a specic demand system, the resulting productivity estimates do change quite drastically if one is no longer ignorant about the product level and the degree of product dierentiation in an industry, and how these factors dier over time and rms. I analyze the impact of trade liberalization on rm performance using the method developed in this paper. Trade liberalization is measured by the abolishment of quota restricted imports and by increased levels of maintained quota. The quota restriction measures serve as additional variables to control for the unobserved price and the resulting estimates for productivity are therefore further puried from demand variation. While I nd positive signicant productivity gains from relaxing quota restrictions, the eects are estimated considerably lower than using standard productivity estimates. The latter still capture price and markup variation (across product segments and time) which are correlated with the change in demand conditions due to a change of trade policy, leading to an overestimation of productivity gains from opening up to trade. 30

Appendix A: The Belgian textile industry and the quota dataset A.1 The Belgian textile industry: I present the structure of the dierent segments, sub-segments and the products in my dataset in Table A.1. The dierent levels are important to structure the regressions and serve as additional sources of variation to identify demand parameters. The number in parentheses indicates the number of subsegments within a given segment whereas the last row indicates the number of products within a given segment. I also estimated demand elasticities at the level of the subsegments, i.e. 52 dierent parameters.

Table A.1.: Segment Structure: Number of Subsegments and Products per Segment
Interior (9) Bed linen Carpets Kitchen linen Matress ticking Table linen Terry toweling articles Trimming Upholstery & furnishing fabr. Wallcoverings 19 Technical (9) Agrotech Buildtech Geotech Indutech Medtech Mobiltech Packtech Protech Sporttech 231 Clothing (18) Fabrics Knitwear Accessories Accessories Baby clothes & childrens Babies wear Mens wear Bath Nightwear & underclothing Childrens wear Others Fabrics for ... Rain-, sportswear & leisure ... Nightwear Womens wear ... Outerwear Workwear & protective suits ... Sportswear Stockings- tights- socks Underwear 61 36 Spinning (9) Blended aramid, polyamid or polyacrylic Blended articial yarns Blended cotton or linen yarns Blended polyester yarns Blended polypropylene or chlorobre yarns Blended yarns Filament Yarns Spun Yarns (> 85% of 1 bre) Synthetic Fibres 84

Finishing (7) Carpeting Knitted fabrics Material before spinning NonWoven Woven fabrics Yarns Specialities

132

31

A.2. Producer prices and demand elasticity As mentioned in the text a producer price index is obtained by taking a weighted average over a representative number of products within an industry, where weights are based on sales (market shares). In the case of Belgium the National Institute of Statistics (NIS) gathers monthly information of market relevant prices (including discounts if available) of around 2,700 representative products (an 8 digit classication - PRODCOM - where the rst 4 are indicating the NACEBELCODE). The index is constructed by using the most recent market share as weights based on sales reported in the ocial tax lings of the relevant companies. The relevant prices take into account both domestic and foreign markets and for some industries both indices are reported. I present unit prices at the 3 digit NACEBELCODE (equivalent to 4/5 digit ISIC code). I constructed these by dividing total value of production in a given subcategory by the quantity produced. Table A2 gives the PPI for the various subcategories with 1994 as base year except for the 175 category (Other textile products, mainly carpets). I do not use these to deate rm-level revenues since I have no information in which category (ies) a rm is active since the product classication cannot be uniquely mapped into the NACEBELCODE and rms are active in various subcategories. The codes have the following description: 171 : Preparation and spinning of textile bres, 172 : Textile weaving, 173 : Finishing of textiles, 174 :.Manufacture of made-up textile articles, except apparel, 175 : Manufacture of other textiles (carpets, ropes, ...), 176 : Manufacture of knitted and crocheted fabrics and 177 : Manufacture of knitted and crocheted articles. Table A.2.: Producer Prices (Unit Prices) at Disaggregated Level 1994 1995 1996 1997 1998 1999 2000 2001 2002 demand elasticity 171 172 100 100 99.4 96.7 100.9 94.5 103.7 94.5 102.8 96.0 95.0 95.8 94.3 94.6 96.7 93.2 97.3 94.2 -5.4675 173 100 110.4 101.1 101.3 108.0 100.6 119.3 108.4 110.7 -3.0628 174 100 111.0 117.9 108.5 117.6 118.2 106.2 107.7 103.1 -3.0628 175 100 99.2 101.5 99.6 102.0 104.1 107.2 n.a. 176 100 100.9 103.4 93.9 93.3 94.8 84.1 86.9 85.8 n.a. 177 100 100.7 94.8 97.5 97.6 92.9 95.5 101.3 106.1 -3.6470

Several observations are important to note. Firstly, there is considerable variation across subcategories of the textiles industry in terms of price changes over the period 1994-2002. The sector Manufacture of knitted and crocheted fabrics (176) has experienced a severe drop in output prices (14.2 percent) over the sample period, whereas the output prices in the Finishing of Textiles (173) has increased with more than 10 percent. Secondly, the evolution in the various subcategories is not smooth, periods of price increases are followed by decreases and the other way around. Thirdly, most of the price decreases occur at the end of the nineties when imports from Central and Eastern Europe were no longer quota restricted as agreed in the Europe Agreements. It is interesting to note that the segment (Spinning) with the most elastic demand (-5.3135) has indeed experienced a negative price evolution (2.7 percent). The latter segment also captures weaving activities which in turn also experienced a price decrease (5.8 percent). The segment (Finishing) with the least elastic demand (-3.2051) has had a sharp increase in its output prices (10.7 percent). The estimated demand elasticities from Table 5 are given in the last row for those subcategories I could map into segments.

32

A.3. Unobserved demand shocks and estimating production function Formally, I relax the assumption that investment only depends on the capital stock and the unobserved productivity shock. I now have two unobservables ( it , it ) and the investment function is now iit = it (kit , it , it ). The demand unobservable it is assumed to follow a Markov process that is independent of the productivity process. We now need a second control sit - say advertisement expenditures - to proxy the unobservable in order to control for the productivity shock. I denote the bivariate policy function determining (iit , sit ) as (.) and assume it is a bijection in (it , it ) conditional on the capital stock kit iit (A.1) = t (kit , it , it ) sit As Ackerberg and Pakes (2005) show this allows us to invert and rewrite the unobservable productivity as a function of the controls in the following way
1 e it = t (kit , iit , sit )

(A.2)

The revenue generating production function is as before and the rst stage of the estimation algorithm now looks as follows
1 r eit = 0 + l lit + m mit + k kit + qIt + t (kit , iit , sit ) + uit e (kit , iit , sit ) + uit = + lit + mit + qIt + 0 l m t

(A.3)

e = kit +1 (kit , iit , sit ). The non parametric function is in three variables, investment, where t k t capital and an additional control, where the latter controls for the unobserved demand shocks it . In addition to the standard Olley and Pakes (1996) methodology I control for both observed and unobserved demand shocks coming from the use of revenue in stead of physical output and from the notion that demand shocks might have an impact on the level of investments. The second stage hardly changes compared to (13) since the process of the demand shock is assumed to be independent of the productivity shock. Consider the revenue generating production function at time t + 1 r eit+1 = 0 + l lit+1 + m mit+1 + k kit+1 + qIt+1 + E ( e it+1 |It ) + it+1 + uit+1

where I have used the fact that productivity and the demand shock follow a rst-order Markov process, i.e. e it+1 = E ( e it+1 | e it ) + it+1 , where it+1 is the news term. The capital coecient is e(.) is dierent compared estimated as before where the only dierence is that the estimate for to the standard case (12) and leads to more precise estimates for the capital stock. b e kit ) + eit+1 r eit+1 bl lit+1 bm mit+1 b qIt+1 = 0 + k kit+1 + g e( it k (A.4)

where eit+1 = it+1 + uit+1 . Variation in output puried from variation in variable inputs and observed demand shock that is correlated with the (observed) control sit is no longer potentially contributed to the variation in capital. In the previous section I collapsed productivity and quality into one unobservable f it . Note that here it implies that I include variables proxying for the quality unobservable (like advertisement expenditures, product dummies as suggested in section 5.2.2.) which take out additional variation related to the demand side, leading to dierent estimates for it in the N LLS estimation. When estimating the capital coecient in equation (A.4) the identifying assumption is that the demand shocks are independent of the productivity shocks. When allowing for productivity to depend on unobserved demand shocks, I would no longer 1 be able to identify the capital coecient as the non parametric function g it k kit , t (kit , iit , sit ) depends on investment at time t. This leaves no more independent variation in the capital stock to identify k as kit+1 = (1 )kit + iit . In fact the only way out is to assume either that this 33

demand unobservable (such as quality) is uncorrelated with capital and ends up in the error term eit+1 . In the Table A.3 I present estimates of product specic markups that are obtained by introducing product dummies to control for unobserved demand shocks in the demand system. I recover product specic estimates and about 20 products are estimated signicantly dierent from their respective segment average. These can be interpreted as product specic Lerner estimates under the identifying assumption that the investment policy function does not depend on unobserved demand shocks. Table A.3 Product specic Lerner estimates (Specication (23)) Segment Product Product Lerner Index
Clothing Rainwear, sportswear and leisure wear: Jackets Rainwear, sportswear and leisure wear: Sportswear Accessories - Labels Textile draining or irrigation Technical sewing thread / Technical weaving Canvas for lm sets and theatre scenery Technical textiles for papermaking industry Textiles for medical care - Hospital linen Upholstery fabrics for car seats Upholstery fabrics for caravans seats (trailers) Special Finishes: Mercerising Special Finishes: Spotrepellent Material before spinning : Cleansing Woven fabrics: Flame retardant Yarns Package dyeing Yarns Sectional warping Yarns Waxing Blended articial yarns CTA/PA Filament Yarns - PA 6 0.4686 0.3132 0.1985 0.7184 0.3458 0.2386 0.4897 0.2432 0.2760 1.4764 1.0276 0.5649 0.6877 1.8124 0.2928 0.3388 0.3829 0.3476 0.3889

Technical

Finishing

Spinning

34

A.4. The quota data The quota data comes straight from the SIGL database constructed by the European Commission (2003) and is publicly available on-line (http://sigl.cec.eu.int/). The quota data is provided using a specic product data classication, the MFA classication. In order to match this to the rm-level data I had to map the MFA classication code into the NACE rev.1 industry code through the PRODCOM classication. I do face the problem that the industry classication is more aggregated than the quota classication which can lead to measurement error in the quota restriction variable. The 182 product categories used in the SIGL database with the relevant unit of measurement (kilograms or units) can be found on-line at http://trade.ec.europa.eu/sigl/products.html. The list of 56 supplying countries facing a quota at some point during the period 19942002 on any of the 182 product categories are: Albania, Argentina, Armenia, Azerbaijan, Bangladesh, Belarus, Bosnia-Herzegovina+Croatia, Brazil, Bulgaria, Cambodia, China, Czech Republic, Egypt, Estonia, Former Yug Rep of Macedonia, Georgia, Hong Kong, Hungary, India, Indonesia, Kazakstan, Kirghistan, Laos, Latvia, Lithuania, Macao, Malaysia, Malta, Moldova, Mongolia, Morocco, Nepal, North Korea, Pakistan, Peru, Philippines, Poland, Romania, Russia, Serbia and Montenegro, Singapore, Slovak Republic, Slovenia, South Korea, Sri Lanka, Syria, Taiwan, Tajikistan, Thailand, Tunisia, Turkey, Turkmenistan, Ukraine, United Arab Emirates, Uzbekistan, Vietnam. Finally, I present two examples that illustrate how the liberalization of trade occurred in the textile industry. I present the evolution of the quota level (level) and the actual ll rate (FR) for two products on imports from China and Poland, respectively.

Table A.4.: Two Examples of Decreased Quota Protection

Example 1 Product Supplier Example 2

Garments other knitted or crocheted Imports from China

Level (x1000, kg) FR (%)

Bed linen, other than knitted or crocheted Imports from Poland

Level (x 1000, kg) FR (%)

Year
1993 1994 1995 1996 1997 1998 1999 2000 2001 2002 2003

21,000 21,630 23,422 24,125 24,848 25,594 26,362 27,153 27,968 30,349 32,932

87.76 99.04 122.85 92.92 103.37 109.00 104.46 99.50 109.81 105.18 105.12

2,600 2,730 3,436 3,787 3,977

60.30 96.19 96.18 89.14 89.41 quota abolished

The table above clearly shows the detailed level of information that is available at each point in time for each product-supplier pair. The liberalization for Bed linen imported from Poland took place under the abolishment of the quota in 1998. Whereas for Garments from China, the increased competition came under the form of increased quota levels (by 88 percent). For both cases, the quota were binding over the span of the period that we study.

35

Appendix B: Production synergies When aggregating the product-level production function to the rm-level, I have assumed that there are no cost synergies or complementarities in producing several products within one rm. However, we know that the textile sector captures both supplying (Spinning and Finishing ) and applying segments (Technical textiles ). Firms that produce both type of products can expect to potentially benet from combining both activities (or more). Therefore, I relax the assumption on the production technology by introducing a parameter sr capturing the complementarity in production of combining dierent products (here segments), where r and s are the dierent segments. More formally the aggregation from product-level production into rm-level is given by (B.1) 5 5 X X m k l Qi = (L M K ) exp( + sr Sisr + uq (B.1) i i) i i i
s=1 r=s

where Sisr is 1 if a rm i is active in both segment r and s and zero otherwise and sr the corresponding coecients. Proceeding as before, I obtain the following augmented production function (B.2).
5 X s=1 5 5 X X s=1 r=s

The estimated segment demand elasticities are somewhat more negative, however, the same economic interpretations apply, i.e. Interior and Spinning are the most elastic segments (-6.81 and -6.76). I now present the estimated coecients on the extra term Sisr in Table B.1.

r eit = 0 + l lit + m mit + k kit + np npit +

s qst Sis +

sr Sisr + it + uit (B.2)

Table B.1: Estimated Product Complementarity sr 1 2 3 4 5 1 -0.37* 2 0.15** -0.27* s 3 0.39* 0.36* -0.61* 4 0.04 0.08 0.28* -0.39* 5 0.35* 0.06 0.23* 0.22* -0.41*

Note: * signi cant at 1% level, **: at 10% level

A positive sign on the coecients in the table above reects a (on average) higher output conditional on inputs and demand conditions for a rm active in any two given segments. Firms combining any activity with Technical textiles (3) generate a higher output. To obtain the entire rm relevant eect, we have to add up the relevant terms, e.g. for a rm active in segment 1 and 3: 0.37 + 0.39 = 0.02, suggesting gains from diversication. The latter is also reected in the negative coecients on the head diagonal. Note that here I only allow test for pair eects in contrast to estimating all potential combinations of segments (31 parameters) .

36

Appendix C. Alternative Proxy Estimators As mentioned in Appendix A of Levinsohn and Petrin (2002) the LP methodology needs rms to operate in a competitive environment and take output and input prices as given in order for the intermediate input to be monotonic increasing in productivity to be able to invert the productivity shocks and proceed as in Olley and Pakes (1996). Models of imperfect competition on the output market do not satisfy those assumptions and the proof depends on the specic degree of competition. Melitz (2001) needs to assume that more productive rms do not set disproportionately higher markups than the less productive rms in order to use the LP procedure. The monotonicity needed in Olley and Pakes (1996) does not depend on the degree of competition on the output market, it just needs the marginal product of capital to be increasing in productivity. I now discuss which additional assumptions one needs in the LP framework in order to allow for non price taking rms. As in LP consider the simple static maximization problem of the rm where the production function is given by Qi = f (Li , mi , i ) where capital is a xed input. The latter is consistent with the OP framework where the capital stock at period t is determined at t 1 through investment and the capital stock. The LP estimator - just like the OP procedure crucially relies upon an invertibility assumption, i.e. demand for intermediate inputs has to be monotonic increasing in productivity. Their proof (Appendix A in Levinsohn and Petrin 2000) works under the assumption of a competitive setting where rms take both input and output prices as given. I now relax this assumption and allow for a more general setting and I show the extra assumption one has to make in order to use the LP approach in setting as discussed in the main text. The prot function of the rm is given by i = pi (Q)Qi pL Li pm mi I now drop the rm index i and the rst order conditions for the inputs labor and materials are given below fL (L, m, ) = pL /p fm (L, m, ) = pm /p and assuming the existence of all second order derivatives, the LP approach works if demand for intermediate inputs are monotonic increasing in the productivity. Dierentiating the FOC with P respect to productivity () and introducing the elasticity of demand = dQ dP Q and < < 0, I obtain the following system L pfLL + fL 2 (pQ ) pfLm + fL fm (pQ ) pfL + fL (pQ )f = m pfm + fm (pQ )f pfmL + fL fm (pQ ) pfmm + fm 2 (pQ ) and we can use Cramers rule to identify the sign of m and establish conditions under which we can still invert the intermediate input demand function, where the sign of the denominator is p 1 always positive since we are working under the maximizing prot condition. Note that pQ = Q which shows the extra assumptions we will need in order for the demand for intermediate inputs to be increasing in the productivity shock m fL f 1 fL fm 1 fL 2 1 fm f 1 sign = sign fL + fmL + fLL + fm + Q Q Q Q Compared to price taking scenario under which LP work, I have four new terms related to the degree of competition ( ). In the case of price taking rms LP need the assumption that fL fmL > fLL fm whereas now we need fL fmL Q + 1 1 (fL fL fm + fL f fmL ) > fLL fm Q + fLL fm f + fL 2 fm 37 (D.2) (D.1)

It is clear that the assumption under the general setting is somewhat more complicated, essen 1). Proceeding with the proof as in LP (2000) since (D.2) tially introducing the markup ( +1 holds everywhere, it holds that m( 2 ; .) > m( 1 ; .) if 2 > 1 The intuition on the extra terms in equation (D.2) is that markups starts playing a role as also noted by Melitz (2001). To see this, consider equation (D.2) and label the terms in the inequality ad follows A + B > C + D. Note that A > C is the sucient assumption needed in the price taking scenario. Furthermore we know that B > 0 and it is generally hard to sign D, the condition (D.1) is now given by fL fmL fLL fm > 1 (D B ) (D.3)

Although the exact conditions are not of interest here, this appendix has shown that relaxing the assumptions of the nature of competition on the output market, has an impact on the validity of the LP estimation algorithm through the invertibility conditions Note that the LP condition is a special case of D.3 where = . As mentioned in the text, recently Ackerberg et al. (2004) analyzed the various proxy estimators used in the literature (OP and LP) and veried how robust they are with respect to the timing of inputs that takes place in the production process. They study the underlying data generating process both proxy estimators assume to identify the production function coecients. Based on their observation I verify whether my estimated demand parameters (markups) are at all sensitive to the underlying assumptions in the production process by using a modied OP estimator. I consider the results discussed in the main text (baseline ) and compare them with the estimated markups obtained from a more exible approach. The exible approach essentially no longer takes a stand on whether all rms face the same factor prices, face unions and more importantly it no longer matters when the productivity shock enters in the timing of the inputs labor and material. The rst stage is then reduced to the following regression rit = qIt + qr qrit + t (lit , mit , kit , iit ) + uit (25)

Table C.1: Estimated Markups under Alternative Specications Specication Baseline Flexible Baseline + Trade protection Flexible + Trade protection Industry 0.31 0.34 0.27 0.22 Interior 0.24 0.12 0.16 0.10 Segment specic Clothing Technical Finishing 0.35 0.31 0.34 0.17 0.17 0.17 0.24 0.21 0.22 0.12 0.13 0.14 Spinning 0.26 0.12 0.19 0.10

Table C.1 above shows, the estimated demand parameters are well within the range of the less exible model used in the main text. Note that the exible approach described in this appendix allows for a general production function where productivity shocks are additive in the log specication and thus allows for exible substitution patterns among inputs (such as the translog production function). However, in order to recover the production function parameters , the similar assumptions used in the main text have to be imposed in the second stage of the Ackerberg et al. (2004) approach. The advantage of the exible approach is that we can estimate the markups in a exible way in the rst stage as robustness check. The last two rows then give us the range of the estimated markups for a given segment, e.g. for the Interior segment the estimated markup lies between 0.16 and 0.10.

38

References
[1] Ackerberg, D,.Caves, K and Frazer, G. 2004. Structural Estimation of Production Functions, U.C.L.A. mimeo [2] Ackerberg, D, Benkard, L., Berry, S. and Pakes, A. 2005. Econometric Tools for Analyzing Market Outcomes, Paper prepared for Handbook of Econometrics, J.J. Heckman ed. [3] Ackerberg, D. and Pakes, A. 2005. Notes on Relaxing the Scalar Unobservable Assumption in Olley and Pakes, mimeo, Harvard University. [4] Arellano M. and Bond, S. 1991. Some tests of specication for panel data: Monte Carlo evidence and an application to employment equations, Review of Economic Studies, vol.58, 277-297. [5] Bartelsman, E. J. and Doms, M. 2000. Understanding Productivity. Lessons from Longitudinal Micro Data, Journal of Economic Literature [6] Bernard, A.B., Redding, S. and Schott, P.K. 2003. Product Choice and Product Switching, NBER WP 9789 [7] Berry, S. 1994. Estimating Discrete-Choice Models of Product Dierentiation, Rand Journal of Economics, 252, 242-262. [8] Berry, S., Levinsohn, L. and Pakes, A. 1995. Automobile Prices in Market Equilibrium, Econometrica, 63, 841-890. [9] Buettner, T. 2004. R&D and the Dynamics of Productivity, mimeo, London School of Economics. [10] Collard-Wexler, A. 2006. Productivity Dispersion and Plant Selection in the Ready-Mix Concrete Industry, mimeo, New York University. [11] De Loecker, J. 2007. Do Exports Generate Higher Productivity? Evidence from Slovenia, Journal of International Economic s, forthcoming. [12] Dunne, T. and Roberts, M.J. 1992. Costs, Demand and Imperfect Competition as Determinants of Plant-Level Output Prices, DP CES 92-5, Center for Economic Studies, US Bureau of the Census. [13] Eslava, M., Haltiwanger, J., Kugler, A. and Kugler, M. 2004. The Eects of Structural Reforms on Productivity and Protability Enhancing Reallocation: Evidence from Colombia, CEPR DP 4569. [14] Ericson, R. and Pakes, A. 1995. Markov Perfect Industry Dynamics: A Framework for Empirical Work, Review of Economic Studies, Vol 62 1, 53-82 [15] European Commission. 2005, website, http://europa.eu.int/comm/enlargement/pas/. [16] Febeltex. 2003.Anual Report 2002-2003 FEBELTEX, Belgium. [17] Foster L., Haltiwanger, J. and Syverson, C. 2005. Reallocation, Firm Turnover and Eciency: Selection on Productivity or Protability?, University of Chicago mimeo. [18] Hall, R.E. 1988.The Relation between Price and Marginal Cost in US Industry, Journal of Political Economy, 96 5, 921-947. [19] Harrison, A.E. 1994. Productivity, Imperfect Competition and Trade Reform: Theory and Evidence, Journal of International Economics, 36, 53-73. [20] Jaumandreu, J. and Mairesse, J. 2004. Using price and demand information to identify production functions, mimeo.

39

[21] Katayama, H., Lu, S. and Tybout, J. 2003. Why Plant-Level Productivity Studies are Often Misleading, and an Alternative Approach to Inference, NBER WP. [22] Klette, T.J. and Griliches, Z 1996. The Inconsistency of Common Scale Estimators When Output Prices are Unobserved and Endogenous, Journal of Applied Econometrics, vol 11 4, 343-361. [23] Konings, J., Van Cayseele P. and Warzynski, F. 2001. The Dynamics of Industrial markups in Two Small Open Economies; Does National Competition Policy Matters?, International Journal of Industrial Organization, 19, 841-859. [24] Konings, J. and Vandenbussche, H. 2005. Antidumping Protection and Markups of Domestic Firms, Journal of International Economics, vol. 65 1, 151 - 165. [25] Levinsohn, J. and Melitz, M. 2002. Productivity in a Dierentiated Products Market Equilibrium, Harvard mimeo. [26] Levinsohn, J. and Petrin, A. 2000. Estimating Production Functions Using Inputs to Control for Unobservables., NBER WP 7819. [27] Levinsohn, J. and Petrin, A. 2003. Estimating Production Functions Using Inputs to Control for Unobservables., Review of Economic Studies, Vol. 70, pp. 317-342. [28] Melitz, M. 2001. Estimating Firm-Level Productivity in Dierentiated Product Industries, Harvard mimeo. [29] Muendler, M.A. 2004. Estimating Production Functions When Productivity Change is Endogenous, UCSD WP 05. [30] Olley, S. and Pakes, A. 1996. The Dynamics of Productivity in the Telecommunications Equipment Industry, Econometrica, Vol 64 6, 1263-98. [31] Pakes, A. 1994, The Estimation of Dynamic Structural Models: Problems and Prospects, Part II. Mixed Continuous-Discrete Control Models and Market Interactions, Chapter 5, pp. 171-259, of Advances in Econometrics: Proceedings of the 6th World Congress of the Econometric Society, edited by J.J. Laont and C. Sims. [32] Pavcnik, N. 2002. Trade Liberalization, Exit, and Productivity Improvement: Evidence from Chilean Plants, Review of Economic Studies, 69 1, 245-76 [33] Petropoulos, W. 2000..Productivity in an industry with dierentiated products, University of Michigan, mimeo. [34] Roeger, W. 1995. Can Imperfect Competition Explain the Dierence between Primal and Dual Productivity Measures? Estimates for US Manufacturing, Journal of Political Economy, 103 2, 316-330. [35] Smarzynska, B. 2004. Does Foreign Direct Investment Increase the Productivity of Domestic Firms? In Search of Spillovers Through Backward Linkages, American Economic Review, 94 3, 605-627. [36] Syverson, C. 2004. Market Structure and Productivity: A Concrete Example., Journal of Political Economy, 112 6, 1181-1222. [37] Tybout, J. 2000. Manufacturing Firms in Developing Countries, Journal of Economic Literature, Vol March, 11-44. [38] Van Biesebroeck, J. 2005. Exporting Raises Productivity in the Sub-Saharan African Manufacturing, Journal of International Economics, 67 (2), 373-391.

40

1 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 1994 1995 clothing 1996 finishing 1997 1998 interior 1999 2000 spinning 2001 technical 2002

Figure 1: Evolution of Quota Protection Measure (qr ) by Segment (1994-2002)

41

Table 1: Summary Statistics of Belgian Textile Industry Year 1994 1995 1996 1997 1998 1999 2000 2001 2002 Average Employment 89 87 83 85 90 88 90 92 99 89 Total Sales 18,412 19,792 18,375 21,561 22,869 21,030 23,698 23,961 26,475 21,828 Value Added 3,940 3,798 3,641 4,365 4,418 4,431 4,617 4,709 5,285 4,367 Capital 2,443 2,378 2,177 2,493 2,650 2,574 2,698 2,679 2,805 2,551 Materials 13,160 14,853 14,313 16,688 17,266 15,546 17,511 17,523 17,053 16,062 PPI 100.00 103.40 99.48 99.17 98.86 98.77 102.98 102.67 102.89

Note: I report averages for all variables in thousands of euro, except for sales where I report total by year.

Table 2: Number of Firms and Production Structure Across Dierent Segments Firms Interior Clothing Technical Finishing Spinning # rms Interior 77.0 Clothing 4.8 58.9 Technical 15.8 33.9 35.1 Finishing 7.3 7.1 19.6 39.6 48 Spinning 1.8 1.8 17.5 12.5 47.5 40

165

56

97

Note: The cells do not have to sum up to 100 percent by row/column, i.e. a rm can be active in more than 2 segments

Table 3: Number of Products and Production Structure Across Dierent Segments Products Interior Clothing Technical Finishing Spinning median min Interior 83.72 3.03 7.01 5.75 3.72 2 1 Clothing 2.78 79.28 8.97 3.52 0.65 6 2 Technical 8.27 15.36 70.16 15.53 27.20 8 1 Finishing 4.41 1.86 9.06 72.85 7.40 11 2 Spinning 0.80 0.48 4.79 2.35 61.04 9 1

Note: The cells do sum up to 100 percent by row. This table has to be read from the rows only.

42

Table 4: The Estimated Coecients of the Production Function OLS [1]

labor materials capital output 0.2300 (0.0095) 0.6298 (0.0074) 0.0879 (0.0072)

KG Level [2]
0.2319 0.2967 (0.0095) (0.0316) 0.6284 0.8041 (0.0074) (0.0770) 0.0868 0.1111 (0.0072) (0.0137) 0.2185 (0.0749) -4.58 1.28 1,291

KG Di [3]
0.2451 0.3338 (0.0198) (0.0343) 0.5958 0.8115 (0.0131) (0.0519) 0.0188 0.0256 (0.0105) (0.0143) 0.2658 (0.0462) -3.76 1.36 1,291

OP [4]
0.2113 (0.0112) 0.6278 (0.0085) 0.0931 (0.0081)

Augmented [5]
0.2126 0.3075 (0.0112) (0.0623) 0.6265 0.9063 (0.0084) (0.1746) 0.1037 0.1500 (0.0063) (0.0337) 0.3087 (0.1335) -3.24 1.45 985

markup Nr Obs 1,291

985

Note: : estimated coecients, : production function coecients. Bootstrapped standard errors are given in parentheses.

Table 5: Estimated Demand Parameters and Implied Firm Elasticities A: Estimated Demand Parameters s No product dummies s ( Sis = 1) s
Interior 0.1888* (0.0742) -5.2966 0.2315* (0.0541) -4.3196 0.2641* -3.7864 0.1673* -5.9773 Clothing 0.2742* (0.1029) -3.6470 0.3140* (0.0756) -3.1847 0.3550* -2.8169 0.2267* -4.4111 Technical 0.2593* (0.0907) -3.8565 0.2952* (0.0648) -3.3875 0.3575* -2.7972 0.2253* -4.4385 Finishing 0.3265* (0.1042) -3.0628 0.3178* (0.0756) -3.1466 0.4563* -2.1915 0.2241* -4.4623 Spinning 0.1829* (0.0774) -5.4675 0.2437* (0.0585) -4.1034 0.2556* -3.9124 0.1455* -6.8729

Product dummies (563 products) s ( Sis = 1) One Segment s (667 obs) s >1 Segments s (318 obs) s B: Implied Firm-Specic Demand Elasticities and markups +1 mean -4.4486 1.3033 s.d. 0.6915 0.0676 minimum -5.4059 1.2269 maximum -3.1627 1.4624

Standard errors are given in parentheses and * denotes signi cance at 1 percent level.

43

Table 6: Number of Quota and Levels in Millions Number of quota protections 1,046 936 824 857 636 642 636 574 486 -54% kg # quota Level 466 3.10 452 3.74 411 3.70 413 3.73 329 4.21 338 4.25 333 4.60 298 5.41 259 5.33 -44% 72% nr of pieces # quota Level 580 8.58 484 9.50 413 7.95 444 9.28 307 9.01 304 10.53 303 9.77 276 11.06 227 12.37 -60% 44%

1994 1995 1996 1997 1998 1999 2000 2001 2002 change

Table 7: The Impact of Additional Demand Information: Quota Restriction Specication (23) without Quota Information with Quota Information 0.2426* 0.1643*
(0.0589) Clothing Markups (0.0658)

Interior

0.3475*
(0.0821)

0.2381*
(0.0915)

Technical Finishing Spinning

0.3126*
(0.0710)

0.2134*
(0.0796)

0.3364*
(0.0824)

0.2219*
(0.0927)

0.2577*
(0.0637)

0.1853*
(0.0690)

qr
quota restriction

-0.0886*
(0.0362)

l m k
returns to scale

0.2514*
(0.0124)

0.2513*
(0.0123)

0.6785*
(0.0100)

0.6808*
(0.0100)

0.0515*
(0.0101)

0.0506*
(0.0122)

[1.30; 1.50]

[1.16; 1.30]

Note: * indicates signicant at 1% and nqr is not signicant

44

Table 8: Impact Trade Liberalization on Productivity Specication (# obs) I (1,291) II (1,088) III (1,088) IV (1,088) Estimated coecient qr qr 4qr 4qr qrt1 V (765) V I (890) 4qr qr level Productivity Estimated using augmented model OP -0.0637** -0.1068*
(0.0366) (0.0296)

-0.0430*
(0.0195)

-0.0612*
(0.0193)

-0.0699*
(0.0312)

-0.1254*
(0.0327)

-0.1172*
(0.0374)

-0.1605*
(0.0393)

-0.0468*
(0.0206)

-0.0348*
(0.0216)

-0.0455**
(0.0272)

-0.1347*
(0.0299)

-0.0584* (0.0229) 0.0019* (0.0008)

-0.0664* (0.0226) -0.0000 (0.0008)

Note: std errors in parentheses, * and ** denote signicant at 5 or lower and 10 percent, resp. All regressions include quota-product classication dummies (23 categories), except for VI.

Table 9: Productivity Impact of a 10 percent Decrease in Protection Measure Productivity Augmented model (1994-1997) (1998-2002) OP Interior 4.37 8.20 2.28 8.06 Clothing 3.60 4.21 3.23 6.45 Technical 4.82 7.32 3.32 8.63 Finishing 1.60 3.05 0.72 2.86 Spinning 4.49 5.71 3.75 8.04 Overall 4.07 6.53 2.59 7.28

Note: The gures are elasticities evaluated at the mean by segment over the relevant period.

45

Table 10: Change in Average Quota Level (1994-2002) Supplying Country Belarus China Hong Kong India Indonesia Malaysia North Korea Pakistan Peru South Korea Taiwan Thailand Uzbekistan Vietnam Products measured in kilograms # pieces 146 60 83 38 62 49 56 127 90 78 58 66 92 129 144 127 61 69 36 28 45 130 556 -92 55

Changes are expressed as a percentage.

46